[RFC PATCH v3 0/3] sched: Skip queued wakeups only when L2 is shared

[RFC PATCH v3 0/3] sched: Skip queued wakeups only when L2 is shared

[Mathieu Desnoyers]

This series improves performance of scheduler wakeups on large systems
by skipping queued wakeups only when CPUs share their L2 cache, rather
than when they share their LLC.

The speedup mainly reproduces on workloads which have at least *some*
idle time (because it significantly increases the number of migrations,
and thus remote wakeups), *and* it needs to have a sufficient load to
cause contention on the runqueue locks.

Feedback is welcome,

Thanks,

Mathieu

Mathieu Desnoyers (3):
  sched: Rename cpus_share_cache to cpus_share_llc
  sched: Introduce cpus_share_l2c (v3)
  sched: ttwu_queue_cond: skip queued wakeups across different l2 caches

Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Julien Desfossez <jdesfossez@digitalocean.com>
Cc: x86@kernel.org

 block/blk-mq.c                 |  2 +-
 include/linux/sched/topology.h | 10 ++++++++--
 kernel/sched/core.c            | 14 +++++++++++---
 kernel/sched/fair.c            |  8 ++++----
 kernel/sched/sched.h           |  2 ++
 kernel/sched/topology.c        | 32 +++++++++++++++++++++++++++++---
 6 files changed, 55 insertions(+), 13 deletions(-)

-- 
2.39.2

[RFC PATCH v3 1/3] sched: Rename cpus_share_cache to cpus_share_llc

[Mathieu Desnoyers]

In preparation for introducing cpus_share_l2c, rename cpus_share_cache
to cpus_share_llc, to make it clear that it specifically groups CPUs by
LLC.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Julien Desfossez <jdesfossez@digitalocean.com>
Cc: x86@kernel.org
---
 block/blk-mq.c                 | 2 +-
 include/linux/sched/topology.h | 4 ++--
 kernel/sched/core.c            | 4 ++--
 kernel/sched/fair.c            | 8 ++++----
 kernel/sched/topology.c        | 2 +-
 5 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index b9f454613989..ed1457ca2c6d 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -1136,7 +1136,7 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 	/* same CPU or cache domain?  Complete locally */
 	if (cpu == rq->mq_ctx->cpu ||
 	    (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
-	     cpus_share_cache(cpu, rq->mq_ctx->cpu)))
+	     cpus_share_llc(cpu, rq->mq_ctx->cpu)))
 		return false;
 
 	/* don't try to IPI to an offline CPU */
diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
index 816df6cc444e..7f9331f71260 100644
--- a/include/linux/sched/topology.h
+++ b/include/linux/sched/topology.h
@@ -178,7 +178,7 @@ extern void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 cpumask_var_t *alloc_sched_domains(unsigned int ndoms);
 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms);
 
-bool cpus_share_cache(int this_cpu, int that_cpu);
+bool cpus_share_llc(int this_cpu, int that_cpu);
 
 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
 typedef int (*sched_domain_flags_f)(void);
@@ -227,7 +227,7 @@ partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 {
 }
 
-static inline bool cpus_share_cache(int this_cpu, int that_cpu)
+static inline bool cpus_share_llc(int this_cpu, int that_cpu)
 {
 	return true;
 }
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index a68d1276bab0..d096ce815099 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3904,7 +3904,7 @@ void wake_up_if_idle(int cpu)
 	rcu_read_unlock();
 }
 
-bool cpus_share_cache(int this_cpu, int that_cpu)
+bool cpus_share_llc(int this_cpu, int that_cpu)
 {
 	if (this_cpu == that_cpu)
 		return true;
@@ -3929,7 +3929,7 @@ static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
 	 * If the CPU does not share cache, then queue the task on the
 	 * remote rqs wakelist to avoid accessing remote data.
 	 */
-	if (!cpus_share_cache(smp_processor_id(), cpu))
+	if (!cpus_share_llc(smp_processor_id(), cpu))
 		return true;
 
 	if (cpu == smp_processor_id())
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 4da5f3541762..680bbe0c7d7a 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6626,7 +6626,7 @@ wake_affine_idle(int this_cpu, int prev_cpu, int sync)
 	 * a cpufreq perspective, it's better to have higher utilisation
 	 * on one CPU.
 	 */
-	if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
+	if (available_idle_cpu(this_cpu) && cpus_share_llc(this_cpu, prev_cpu))
 		return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
 
 	if (sync && cpu_rq(this_cpu)->nr_running == 1)
@@ -7146,7 +7146,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	/*
 	 * If the previous CPU is cache affine and idle, don't be stupid:
 	 */
-	if (prev != target && cpus_share_cache(prev, target) &&
+	if (prev != target && cpus_share_llc(prev, target) &&
 	    (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
 	    asym_fits_cpu(task_util, util_min, util_max, prev))
 		return prev;
@@ -7172,7 +7172,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	p->recent_used_cpu = prev;
 	if (recent_used_cpu != prev &&
 	    recent_used_cpu != target &&
-	    cpus_share_cache(recent_used_cpu, target) &&
+	    cpus_share_llc(recent_used_cpu, target) &&
 	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
 	    cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
 	    asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
@@ -7206,7 +7206,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	if (sched_smt_active()) {
 		has_idle_core = test_idle_cores(target);
 
-		if (!has_idle_core && cpus_share_cache(prev, target)) {
+		if (!has_idle_core && cpus_share_llc(prev, target)) {
 			i = select_idle_smt(p, prev);
 			if ((unsigned int)i < nr_cpumask_bits)
 				return i;
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 6682535e37c8..1ae2a0a1115a 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -661,7 +661,7 @@ static void destroy_sched_domains(struct sched_domain *sd)
  *
  * Also keep a unique ID per domain (we use the first CPU number in
  * the cpumask of the domain), this allows us to quickly tell if
- * two CPUs are in the same cache domain, see cpus_share_cache().
+ * two CPUs are in the same cache domain, see cpus_share_llc().
  */
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
-- 
2.39.2

[RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

Introduce cpus_share_l2c to allow querying whether two logical CPUs
share a common L2 cache.

Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
cache has a latency of 4-5 cycles, the L2 cache has a latency of at
least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
this, I measured the RAM accesses to a latency around 120ns on my
system [2]. So L3 really is only 2.4x faster than RAM accesses.
Therefore, with this relatively slow access speed compared to L2, the
scheduler will benefit from only considering CPUs sharing an L2 cache
for the purpose of using remote runqueue locking rather than queued
wakeups.

Link: https://en.wikichip.org/wiki/amd/microarchitectures/zen_4 [1]
Link: https://github.com/ChipsandCheese/MemoryLatencyTest [2]
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Julien Desfossez <jdesfossez@digitalocean.com>
Cc: x86@kernel.org
---
Changes since v1:
- Fix l2c id for configurations where L2 have a single logical CPU:
  use TOPOLOGY_CLUSTER_SYSFS to find out whether topology cluster is
  implemented or if LLC should be used as fallback.

Changes since v2:
- Reverse order of cpu_get_l2c_info() l2c_id and l2c_size output
  arguments to match the caller.
---
 include/linux/sched/topology.h |  6 ++++++
 kernel/sched/core.c            |  8 ++++++++
 kernel/sched/sched.h           |  2 ++
 kernel/sched/topology.c        | 32 +++++++++++++++++++++++++++++---
 4 files changed, 45 insertions(+), 3 deletions(-)

diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
index 7f9331f71260..c5fdee188bea 100644
--- a/include/linux/sched/topology.h
+++ b/include/linux/sched/topology.h
@@ -178,6 +178,7 @@ extern void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 cpumask_var_t *alloc_sched_domains(unsigned int ndoms);
 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms);
 
+bool cpus_share_l2c(int this_cpu, int that_cpu);
 bool cpus_share_llc(int this_cpu, int that_cpu);
 
 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
@@ -227,6 +228,11 @@ partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 {
 }
 
+static inline bool cpus_share_l2c(int this_cpu, int that_cpu)
+{
+	return true;
+}
+
 static inline bool cpus_share_llc(int this_cpu, int that_cpu)
 {
 	return true;
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index d096ce815099..11e60a69ae31 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3904,6 +3904,14 @@ void wake_up_if_idle(int cpu)
 	rcu_read_unlock();
 }
 
+bool cpus_share_l2c(int this_cpu, int that_cpu)
+{
+	if (this_cpu == that_cpu)
+		return true;
+
+	return per_cpu(sd_l2c_id, this_cpu) == per_cpu(sd_l2c_id, that_cpu);
+}
+
 bool cpus_share_llc(int this_cpu, int that_cpu)
 {
 	if (this_cpu == that_cpu)
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 81ac605b9cd5..d93543db214c 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1828,6 +1828,8 @@ static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
 	return sd;
 }
 
+DECLARE_PER_CPU(int, sd_l2c_size);
+DECLARE_PER_CPU(int, sd_l2c_id);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 DECLARE_PER_CPU(int, sd_llc_size);
 DECLARE_PER_CPU(int, sd_llc_id);
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 1ae2a0a1115a..fadb66edcf5e 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -661,8 +661,11 @@ static void destroy_sched_domains(struct sched_domain *sd)
  *
  * Also keep a unique ID per domain (we use the first CPU number in
  * the cpumask of the domain), this allows us to quickly tell if
- * two CPUs are in the same cache domain, see cpus_share_llc().
+ * two CPUs are in the same cache domain, see cpus_share_l2c() and
+ * cpus_share_llc().
  */
+DEFINE_PER_CPU(int, sd_l2c_size);
+DEFINE_PER_CPU(int, sd_l2c_id);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
 DEFINE_PER_CPU(int, sd_llc_id);
@@ -672,12 +675,27 @@ DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 
+#ifdef TOPOLOGY_CLUSTER_SYSFS
+static int cpu_get_l2c_info(int cpu, int *l2c_size, int *l2c_id)
+{
+	const struct cpumask *cluster_mask = topology_cluster_cpumask(cpu);
+
+	*l2c_size = cpumask_weight(cluster_mask);
+	*l2c_id = cpumask_first(cluster_mask);
+	return 0;
+}
+#else
+static int cpu_get_l2c_info(int cpu, int *l2c_size, int *l2c_id)
+{
+	return -1;
+}
+#endif
+
 static void update_top_cache_domain(int cpu)
 {
 	struct sched_domain_shared *sds = NULL;
 	struct sched_domain *sd;
-	int id = cpu;
-	int size = 1;
+	int id = cpu, size = 1, l2c_id, l2c_size;
 
 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 	if (sd) {
@@ -686,6 +704,14 @@ static void update_top_cache_domain(int cpu)
 		sds = sd->shared;
 	}
 
+	if (cpu_get_l2c_info(cpu, &l2c_size, &l2c_id)) {
+		/* Fallback on using LLC. */
+		l2c_size = size;
+		l2c_id = id;
+	}
+	per_cpu(sd_l2c_size, cpu) = l2c_size;
+	per_cpu(sd_l2c_id, cpu) = l2c_id;
+
 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 	per_cpu(sd_llc_size, cpu) = size;
 	per_cpu(sd_llc_id, cpu) = id;
-- 
2.39.2

[RFC PATCH v3 3/3] sched: ttwu_queue_cond: skip queued wakeups across different l2 caches

[Mathieu Desnoyers]

Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
cache has a latency of 4-5 cycles, the L2 cache has a latency of at
least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
this, I measured the RAM accesses to a latency around 120ns on my
system [2]. So L3 really is only 2.4x faster than RAM accesses.
Therefore, with this relatively slow access speed compared to L2, the
scheduler will benefit from only considering CPUs sharing an L2 cache
for the purpose of using remote runqueue locking rather than queued
wakeups.

Skipping queued wakeups for all logical CPUs sharing an LLC means that
on a 192 cores AMD EPYC 9654 96-Core Processor (over 2 sockets), groups
of 8 cores (16 hardware threads) end up grabbing runqueue locks of other
runqueues within the same group for each wakeup, causing contention on
the runqueue locks.

Improve this by only considering logical cpus sharing an L2 cache as
candidates for skipping use of the queued wakeups.

This results in the following benchmark improvements:

    hackbench -g 32 -f 20 --threads --pipe -l 480000 -s 100

from 49s to 34s. (30% speedup)

And similarly with perf bench:

    perf bench sched messaging -g 32 -p -t -l 100000

from 10.9s to 7.4s (32% speedup)

I have noticed that in order to observe the speedup, the workload needs
to keep the CPUs sufficiently busy to cause runqueue lock contention,
but not so busy that they don't go idle. This can be explained by the
fact that idle CPUs are a preferred target for task wakeup runqueue
selection, and therefore having idle cpus causes more migrations, which
triggers more remote wakeups. For both the hackbench and the perf bench
sched messaging benchmarks, the scale of the workload can be tweaked by
changing the number groups.

This was developed as part of the investigation into a weird regression
reported by AMD where adding a raw spinlock in the scheduler context
switch accelerated hackbench. It turned out that changing this raw
spinlock for a loop of 10000x cpu_relax within do_idle() had similar
benefits.

This patch achieves a similar effect without busy waiting nor changing
anything about runqueue selection on wakeup. It considers that only
hardware threads sharing an L2 cache should skip the queued
try-to-wakeup and directly grab the target runqueue lock, rather than
allowing all hardware threads sharing an LLC to do so.

I would be interested to hear feedback about performance impact of this
patch (improvement or regression) on other workloads and hardware,
especially for Intel CPUs. One thing that we might want to empirically
figure out from the topology is whether there is a maximum number of
hardware threads within an LLC below which it would make sense to use
the LLC rather than L2 as group within which queued wakeups can be
skipped.

Link: https://en.wikichip.org/wiki/amd/microarchitectures/zen_4 [1]
Link: https://github.com/ChipsandCheese/MemoryLatencyTest [2]
Link: https://lore.kernel.org/r/09e0f469-a3f7-62ef-75a1-e64cec2dcfc5@amd.com
Link: https://lore.kernel.org/lkml/20230725193048.124796-1-mathieu.desnoyers@efficios.com/
Link: https://lore.kernel.org/lkml/20230810140635.75296-1-mathieu.desnoyers@efficios.com/
Link: https://lore.kernel.org/lkml/20230810140635.75296-1-mathieu.desnoyers@efficios.com/
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Julien Desfossez <jdesfossez@digitalocean.com>
Cc: x86@kernel.org
---
 kernel/sched/core.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 11e60a69ae31..317f4cec4653 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3934,10 +3934,10 @@ static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
 		return false;
 
 	/*
-	 * If the CPU does not share cache, then queue the task on the
+	 * If the CPU does not share L2 cache, then queue the task on the
 	 * remote rqs wakelist to avoid accessing remote data.
 	 */
-	if (!cpus_share_llc(smp_processor_id(), cpu))
+	if (!cpus_share_l2c(smp_processor_id(), cpu))
 		return true;
 
 	if (cpu == smp_processor_id())
-- 
2.39.2

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

On 8/22/23 07:31, Mathieu Desnoyers wrote:
> Introduce cpus_share_l2c to allow querying whether two logical CPUs
> share a common L2 cache.
> 
> Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
> cache has a latency of 4-5 cycles, the L2 cache has a latency of at
> least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
> this, I measured the RAM accesses to a latency around 120ns on my
> system [2]. So L3 really is only 2.4x faster than RAM accesses.
> Therefore, with this relatively slow access speed compared to L2, the
> scheduler will benefit from only considering CPUs sharing an L2 cache
> for the purpose of using remote runqueue locking rather than queued
> wakeups.

So I did some more benchmarking to figure out whether the reason for 
this speedup is the latency delta between L2 and L3, or is due to the 
number of hw threads contending on the rq locks.

I tried to force grouping of those "skip ttwu queue" groups by a subset 
of the LLC id, basically by taking the LLC id and adding the cpu number 
modulo N, where N is chosen based on my machine topology.

The end result is that I have similar numbers for groups of 1, 2, 4 HW 
threads (which use rq locks and skip queued ttwu within the group). 
Starting with group of size 8, the performance starts to degrade.

So I wonder: do machines with more than 4 HW threads per L2 cache exist? 
If it's the case, there we should think about grouping not only by L2 
cache, but also sub-divide this group so the number of hw threads per 
group is at most 4.

Here are my results with the hackbench test-case:

Group cpus by 16 hw threads:

Time: 49s

- group cpus by 8 hw threads: (llc_id + cpu modulo 2)

Time: 39s

- group cpus by 4 hw threads: (llc_id + cpu modulo 4)

Time: 34s

- group cpus by 2 hw threads: (llc_id + cpu modulo 8)
(expect same as L2 grouping on this machine)

Time: 34s

- group cpus by 1 hw threads: (cpu)

Time: 33s

Thanks,

Mathieu



> 
> Link: https://en.wikichip.org/wiki/amd/microarchitectures/zen_4 [1]
> Link: https://github.com/ChipsandCheese/MemoryLatencyTest [2]
> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Valentin Schneider <vschneid@redhat.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Ben Segall <bsegall@google.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
> Cc: Aaron Lu <aaron.lu@intel.com>
> Cc: Julien Desfossez <jdesfossez@digitalocean.com>
> Cc: x86@kernel.org
> ---
> Changes since v1:
> - Fix l2c id for configurations where L2 have a single logical CPU:
>    use TOPOLOGY_CLUSTER_SYSFS to find out whether topology cluster is
>    implemented or if LLC should be used as fallback.
> 
> Changes since v2:
> - Reverse order of cpu_get_l2c_info() l2c_id and l2c_size output
>    arguments to match the caller.
> ---
>   include/linux/sched/topology.h |  6 ++++++
>   kernel/sched/core.c            |  8 ++++++++
>   kernel/sched/sched.h           |  2 ++
>   kernel/sched/topology.c        | 32 +++++++++++++++++++++++++++++---
>   4 files changed, 45 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
> index 7f9331f71260..c5fdee188bea 100644
> --- a/include/linux/sched/topology.h
> +++ b/include/linux/sched/topology.h
> @@ -178,6 +178,7 @@ extern void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
>   cpumask_var_t *alloc_sched_domains(unsigned int ndoms);
>   void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms);
>   
> +bool cpus_share_l2c(int this_cpu, int that_cpu);
>   bool cpus_share_llc(int this_cpu, int that_cpu);
>   
>   typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
> @@ -227,6 +228,11 @@ partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
>   {
>   }
>   
> +static inline bool cpus_share_l2c(int this_cpu, int that_cpu)
> +{
> +	return true;
> +}
> +
>   static inline bool cpus_share_llc(int this_cpu, int that_cpu)
>   {
>   	return true;
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index d096ce815099..11e60a69ae31 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3904,6 +3904,14 @@ void wake_up_if_idle(int cpu)
>   	rcu_read_unlock();
>   }
>   
> +bool cpus_share_l2c(int this_cpu, int that_cpu)
> +{
> +	if (this_cpu == that_cpu)
> +		return true;
> +
> +	return per_cpu(sd_l2c_id, this_cpu) == per_cpu(sd_l2c_id, that_cpu);
> +}
> +
>   bool cpus_share_llc(int this_cpu, int that_cpu)
>   {
>   	if (this_cpu == that_cpu)
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index 81ac605b9cd5..d93543db214c 100644
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -1828,6 +1828,8 @@ static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
>   	return sd;
>   }
>   
> +DECLARE_PER_CPU(int, sd_l2c_size);
> +DECLARE_PER_CPU(int, sd_l2c_id);
>   DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
>   DECLARE_PER_CPU(int, sd_llc_size);
>   DECLARE_PER_CPU(int, sd_llc_id);
> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> index 1ae2a0a1115a..fadb66edcf5e 100644
> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -661,8 +661,11 @@ static void destroy_sched_domains(struct sched_domain *sd)
>    *
>    * Also keep a unique ID per domain (we use the first CPU number in
>    * the cpumask of the domain), this allows us to quickly tell if
> - * two CPUs are in the same cache domain, see cpus_share_llc().
> + * two CPUs are in the same cache domain, see cpus_share_l2c() and
> + * cpus_share_llc().
>    */
> +DEFINE_PER_CPU(int, sd_l2c_size);
> +DEFINE_PER_CPU(int, sd_l2c_id);
>   DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
>   DEFINE_PER_CPU(int, sd_llc_size);
>   DEFINE_PER_CPU(int, sd_llc_id);
> @@ -672,12 +675,27 @@ DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
>   DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
>   DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
>   
> +#ifdef TOPOLOGY_CLUSTER_SYSFS
> +static int cpu_get_l2c_info(int cpu, int *l2c_size, int *l2c_id)
> +{
> +	const struct cpumask *cluster_mask = topology_cluster_cpumask(cpu);
> +
> +	*l2c_size = cpumask_weight(cluster_mask);
> +	*l2c_id = cpumask_first(cluster_mask);
> +	return 0;
> +}
> +#else
> +static int cpu_get_l2c_info(int cpu, int *l2c_size, int *l2c_id)
> +{
> +	return -1;
> +}
> +#endif
> +
>   static void update_top_cache_domain(int cpu)
>   {
>   	struct sched_domain_shared *sds = NULL;
>   	struct sched_domain *sd;
> -	int id = cpu;
> -	int size = 1;
> +	int id = cpu, size = 1, l2c_id, l2c_size;
>   
>   	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
>   	if (sd) {
> @@ -686,6 +704,14 @@ static void update_top_cache_domain(int cpu)
>   		sds = sd->shared;
>   	}
>   
> +	if (cpu_get_l2c_info(cpu, &l2c_size, &l2c_id)) {
> +		/* Fallback on using LLC. */
> +		l2c_size = size;
> +		l2c_id = id;
> +	}
> +	per_cpu(sd_l2c_size, cpu) = l2c_size;
> +	per_cpu(sd_l2c_id, cpu) = l2c_id;
> +
>   	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
>   	per_cpu(sd_llc_size, cpu) = size;
>   	per_cpu(sd_llc_id, cpu) = id;

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

On 8/23/23 11:26, Mathieu Desnoyers wrote:
> On 8/22/23 07:31, Mathieu Desnoyers wrote:
>> Introduce cpus_share_l2c to allow querying whether two logical CPUs
>> share a common L2 cache.
>>
>> Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
>> cache has a latency of 4-5 cycles, the L2 cache has a latency of at
>> least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
>> this, I measured the RAM accesses to a latency around 120ns on my
>> system [2]. So L3 really is only 2.4x faster than RAM accesses.
>> Therefore, with this relatively slow access speed compared to L2, the
>> scheduler will benefit from only considering CPUs sharing an L2 cache
>> for the purpose of using remote runqueue locking rather than queued
>> wakeups.
> 
> So I did some more benchmarking to figure out whether the reason for 
> this speedup is the latency delta between L2 and L3, or is due to the 
> number of hw threads contending on the rq locks.
> 
> I tried to force grouping of those "skip ttwu queue" groups by a subset 
> of the LLC id, basically by taking the LLC id and adding the cpu number 
> modulo N, where N is chosen based on my machine topology.
> 
> The end result is that I have similar numbers for groups of 1, 2, 4 HW 
> threads (which use rq locks and skip queued ttwu within the group). 
> Starting with group of size 8, the performance starts to degrade.
> 
> So I wonder: do machines with more than 4 HW threads per L2 cache exist? 
> If it's the case, there we should think about grouping not only by L2 
> cache, but also sub-divide this group so the number of hw threads per 
> group is at most 4.
> 
> Here are my results with the hackbench test-case:
> 
> Group cpus by 16 hw threads:
> 
> Time: 49s
> 
> - group cpus by 8 hw threads: (llc_id + cpu modulo 2)
> 
> Time: 39s
> 
> - group cpus by 4 hw threads: (llc_id + cpu modulo 4)
> 
> Time: 34s
> 
> - group cpus by 2 hw threads: (llc_id + cpu modulo 8)
> (expect same as L2 grouping on this machine)
> 
> Time: 34s
> 
> - group cpus by 1 hw threads: (cpu)
> 
> Time: 33s

One more interesting data point: I tried modifying the grouping
so that I would explicitly group by hw threads which sit in different
L3, and even on different NUMA nodes for some
(group id = cpu_id % 192). This is expected to generate really _bad_
cache locality for the runqueue locks within a group.

The result for these groups of 3 HW threads is about 33s with the
hackbench benchmark, which seems to confirm that the cause of the
speedup is reduction of the contention on the rq locks by making the
groups smaller, and therefore reducing the likelihood of contention for 
the rq locks, rather than by improving cache locality from L3 to L2.

So grouping by shared L2 only happens to make the group size OK, but
this benchmark does not significantly benefit from having all runqueue
locks on the same L2.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Wed, Aug 23, 2023 at 02:52:17PM -0400, Mathieu Desnoyers wrote:
> On 8/23/23 11:26, Mathieu Desnoyers wrote:
> > On 8/22/23 07:31, Mathieu Desnoyers wrote:
> > > Introduce cpus_share_l2c to allow querying whether two logical CPUs
> > > share a common L2 cache.
> > > 
> > > Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
> > > cache has a latency of 4-5 cycles, the L2 cache has a latency of at
> > > least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
> > > this, I measured the RAM accesses to a latency around 120ns on my
> > > system [2]. So L3 really is only 2.4x faster than RAM accesses.
> > > Therefore, with this relatively slow access speed compared to L2, the
> > > scheduler will benefit from only considering CPUs sharing an L2 cache
> > > for the purpose of using remote runqueue locking rather than queued
> > > wakeups.
> > 
> > So I did some more benchmarking to figure out whether the reason for
> > this speedup is the latency delta between L2 and L3, or is due to the
> > number of hw threads contending on the rq locks.
> > 
> > I tried to force grouping of those "skip ttwu queue" groups by a subset
> > of the LLC id, basically by taking the LLC id and adding the cpu number
> > modulo N, where N is chosen based on my machine topology.
> > 
> > The end result is that I have similar numbers for groups of 1, 2, 4 HW
> > threads (which use rq locks and skip queued ttwu within the group).
> > Starting with group of size 8, the performance starts to degrade.
> > 
> > So I wonder: do machines with more than 4 HW threads per L2 cache exist?
> > If it's the case, there we should think about grouping not only by L2
> > cache, but also sub-divide this group so the number of hw threads per
> > group is at most 4.
> > 
> > Here are my results with the hackbench test-case:
> > 
> > Group cpus by 16 hw threads:
> > 
> > Time: 49s
> > 
> > - group cpus by 8 hw threads: (llc_id + cpu modulo 2)
> > 
> > Time: 39s
> > 
> > - group cpus by 4 hw threads: (llc_id + cpu modulo 4)
> > 
> > Time: 34s
> > 
> > - group cpus by 2 hw threads: (llc_id + cpu modulo 8)
> > (expect same as L2 grouping on this machine)
> > 
> > Time: 34s
> > 
> > - group cpus by 1 hw threads: (cpu)
> > 
> > Time: 33s
> 
> One more interesting data point: I tried modifying the grouping
> so that I would explicitly group by hw threads which sit in different
> L3, and even on different NUMA nodes for some
> (group id = cpu_id % 192). This is expected to generate really _bad_
> cache locality for the runqueue locks within a group.
> 
> The result for these groups of 3 HW threads is about 33s with the
> hackbench benchmark, which seems to confirm that the cause of the
> speedup is reduction of the contention on the rq locks by making the
> groups smaller, and therefore reducing the likelihood of contention for the
> rq locks, rather than by improving cache locality from L3 to L2.

In addition to reduced rq lock contention, I think another reason this
improves performance is because it reduced task migration. Not sure if
it is the case on your test system, but on my test machine(Intel SPR),
task migration number dropped.

Hackbench on Intel SPR(2sockets/112cores/224threads) test summary:
- performance improved for all three cases; the more tasks(groups), the
  more performance gain;
- task migrations dropped with this series for nr_group=20 and 32
  according to 'perf stat'. migration number didn't drop for nr_group=10
  but the two update functions' cost dropped which means fewer access to
  tg->load_avg and thus, fewer task migrations. This is contradictory
  and I can not explain yet;
- rq lock contention dropped for all three cases and it dropped the most
  under more overloaded case: nr_group=32.

It's not clear to me why this series can reduce task migrations. I doubt
it has something to do with more wakelist style wakeup becasue for this
test machine, only a single core with two SMT threads share L2 so more
wakeups are through wakelist. In wakelist style wakeup, the target rq's
ttwu_pending is set and that will make the target cpu as !idle_cpu();
This is faster than grabbing the target rq's lock and then increase
target rq's nr_running or set target rq's curr to something else than
idle. So wakelist style wakeup can make target cpu appear as non idle
faster, but I can't connect this with reduced migration yet, I just feel
this might be the reason why task migration reduced.

Below are detailed test data.
Base: 6.5-rc1.
rq_spin%: The percent of raw_spin_rq_lock_nested() as reported by
          perf/events=cycles:pp
migration: cpu-migrations reported by "perf stat -a -- sleep 5"

The cmdline used is:
hackbench -g $nr_group -f 20 --pipe --threads -l 480000 -s 100

nr_group=10:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base         46s    1.32%        20.06%             10.78%      10.227 K/sec
this_series  37s    0.57%        15.08%              7.05%      10.722 K/sec

nr_group=20:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base         69s    2.57%        19.68%             10.74%      12.098 K/sec
this_series  41s    0.62%        12.11%              5.78%       8.617 K/sec

nr_group=32:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base     192s±25%  15.12%        25.83%             9.33%       12.141 K/sec
this_series  71s    0.47%        10.98%             4.58%        8.198 K/sec

I also tested applying my "ratelimit update to tg->load_avg" patch and
the test summary is:
- performance improved noticeably for nr_group=20 and slightly for
  nr_group=10 case; nr_group=32's performance is roughly the same.
- task migrations dropped for all three cases; nr_group=20 saw the
  biggest drop.
- rq lock contention dropped for all three cases and again, nr_group=32
  saw the biggest drop.

Below are detailed data.
Base: peter's sched/core branch with my "ratelimit" patch.
this_series: apply this patchset on top of base.

nr_group=10:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base         36s    0.55%        0.46%              1.43%       15.034 K/sec
this_series  35s    0.56%        0.52%              1.53%       13.751 K/sec

nr_group=20:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base         47s    1.28%        0.73%              2.33%       21.217 K/sec
this_series  42s    0.60%        0.69%              1.69%       14.130 K/sec

nr_group=32:
            time  rq_spin%  update_cfs_group%  update_load_avg% migration
base         70s    2.38%        0.60%              2.19%       17.855 K/sec
this_series  70s    0.58%        0.63%              1.77%       12.331 K/sec

Thanks,
Aaron

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

On 8/24/23 03:52, Aaron Lu wrote:
> On Wed, Aug 23, 2023 at 02:52:17PM -0400, Mathieu Desnoyers wrote:
>> On 8/23/23 11:26, Mathieu Desnoyers wrote:
>>> On 8/22/23 07:31, Mathieu Desnoyers wrote:
>>>> Introduce cpus_share_l2c to allow querying whether two logical CPUs
>>>> share a common L2 cache.
>>>>
>>>> Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
>>>> cache has a latency of 4-5 cycles, the L2 cache has a latency of at
>>>> least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
>>>> this, I measured the RAM accesses to a latency around 120ns on my
>>>> system [2]. So L3 really is only 2.4x faster than RAM accesses.
>>>> Therefore, with this relatively slow access speed compared to L2, the
>>>> scheduler will benefit from only considering CPUs sharing an L2 cache
>>>> for the purpose of using remote runqueue locking rather than queued
>>>> wakeups.
>>>
>>> So I did some more benchmarking to figure out whether the reason for
>>> this speedup is the latency delta between L2 and L3, or is due to the
>>> number of hw threads contending on the rq locks.
>>>
>>> I tried to force grouping of those "skip ttwu queue" groups by a subset
>>> of the LLC id, basically by taking the LLC id and adding the cpu number
>>> modulo N, where N is chosen based on my machine topology.
>>>
>>> The end result is that I have similar numbers for groups of 1, 2, 4 HW
>>> threads (which use rq locks and skip queued ttwu within the group).
>>> Starting with group of size 8, the performance starts to degrade.
>>>
>>> So I wonder: do machines with more than 4 HW threads per L2 cache exist?
>>> If it's the case, there we should think about grouping not only by L2
>>> cache, but also sub-divide this group so the number of hw threads per
>>> group is at most 4.
>>>
>>> Here are my results with the hackbench test-case:
>>>
>>> Group cpus by 16 hw threads:
>>>
>>> Time: 49s
>>>
>>> - group cpus by 8 hw threads: (llc_id + cpu modulo 2)
>>>
>>> Time: 39s
>>>
>>> - group cpus by 4 hw threads: (llc_id + cpu modulo 4)
>>>
>>> Time: 34s
>>>
>>> - group cpus by 2 hw threads: (llc_id + cpu modulo 8)
>>> (expect same as L2 grouping on this machine)
>>>
>>> Time: 34s
>>>
>>> - group cpus by 1 hw threads: (cpu)
>>>
>>> Time: 33s
>>
>> One more interesting data point: I tried modifying the grouping
>> so that I would explicitly group by hw threads which sit in different
>> L3, and even on different NUMA nodes for some
>> (group id = cpu_id % 192). This is expected to generate really _bad_
>> cache locality for the runqueue locks within a group.
>>
>> The result for these groups of 3 HW threads is about 33s with the
>> hackbench benchmark, which seems to confirm that the cause of the
>> speedup is reduction of the contention on the rq locks by making the
>> groups smaller, and therefore reducing the likelihood of contention for the
>> rq locks, rather than by improving cache locality from L3 to L2.
> 
> In addition to reduced rq lock contention, I think another reason this
> improves performance is because it reduced task migration. Not sure if
> it is the case on your test system, but on my test machine(Intel SPR),
> task migration number dropped.

Yes, it's indeed the case on my system as well. It cuts migrations by 
half (9.2K/sec down to 5.0K/sec).

> Hackbench on Intel SPR(2sockets/112cores/224threads) test summary:
> - performance improved for all three cases; the more tasks(groups), the
>    more performance gain;

Interesting!

> - task migrations dropped with this series for nr_group=20 and 32
>    according to 'perf stat'. migration number didn't drop for nr_group=10
>    but the two update functions' cost dropped which means fewer access to
>    tg->load_avg and thus, fewer task migrations. This is contradictory
>    and I can not explain yet;

Neither can I.

> - rq lock contention dropped for all three cases and it dropped the most
>    under more overloaded case: nr_group=32.

The fact that you observed rq lock contention dropping is a good sign
that doing more queued wakeups is a good thing compared to allowing
non-queued wakeups across cpus sharing a whole LLC.

At this point I'm not sure if the reduction on rq lock contention is 
mostly due to using queued wakeups rather than grabbing remote rq locks, 
or by a side-effet of doing a queued wakeup rather than immediately 
doing the wakeup, which would open a window where the target rq is still 
considered idle by the various code paths within select_task_rq_fair 
which don't care about rq->ttwu_pending.

> It's not clear to me why this series can reduce task migrations. I doubt
> it has something to do with more wakelist style wakeup becasue for this
> test machine, only a single core with two SMT threads share L2 so more
> wakeups are through wakelist. In wakelist style wakeup, the target rq's
> ttwu_pending is set and that will make the target cpu as !idle_cpu();
> This is faster than grabbing the target rq's lock and then increase
> target rq's nr_running or set target rq's curr to something else than
> idle. So wakelist style wakeup can make target cpu appear as non idle
> faster, but I can't connect this with reduced migration yet, I just feel
> this might be the reason why task migration reduced.

Many code paths in select_task_rq_fair don't seem to care about 
rq->ttwu_pending.

In wake_affine_idle, for sync wakeups, if nr_running is 1 on the waker, 
we choose the waker cpu as target.

In wake_affine_idle, if none of waker or prev wakee cpus are idle, then 
it uses wake_affine_weight to find out which of the waker/prev wakee 
cpus are targets based on their respective load.

If wake_affine_idle cannot find an idle waker/prev wakee cpu, and if 
wake_affine_weight finds that the prev wakee cpu had a lower load, then
wake_affine returns the prev wakee cpu as target. This happens even if 
the prev wakee cpu is not idle.

This "target" cpu is then passed to select_idle_sibling. It expects the 
available_idle_cpu(target) to check again to see whether the target cpu 
is idle. However, it also uses "sched_idle_cpu(target)" which _only_ 
considers nr_running (not ttwu_pending flag). Likewise for the other 
similar idleness checks below in select_idle_sibling for prev and 
recent_used_cpu. The same happens for the case where a per-cpu kthread
stacks with the wakee.

I've tried adding checks for rq->ttwu_pending in those code paths on top 
of my patch and I'm still observing the reduction in number of 
migrations, so it's unclear to me how doing more queued wakeups can 
reduce migrations the way it does.

I'm starting to think may want to explore explicitly rate limiting task 
migrations as well.

For instance, we could do something like this:

Within a 1ms window, if a task is migrated more than once, the following 
wakeups would consider that the prev runqueue should be considered in 
priority (as if it was completely idle) as long as its load is below a 
given threshold.

So every 1ms tasks have a chance to be migrated to the idlest runqueues, 
but we would then eliminate those frequent migration patterns which end 
up being bad for cache locality.

Thoughts ?

Thanks,

Mathieu


> 
> Below are detailed test data.
> Base: 6.5-rc1.
> rq_spin%: The percent of raw_spin_rq_lock_nested() as reported by
>            perf/events=cycles:pp
> migration: cpu-migrations reported by "perf stat -a -- sleep 5"
> 
> The cmdline used is:
> hackbench -g $nr_group -f 20 --pipe --threads -l 480000 -s 100
> 
> nr_group=10:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base         46s    1.32%        20.06%             10.78%      10.227 K/sec
> this_series  37s    0.57%        15.08%              7.05%      10.722 K/sec
> 
> nr_group=20:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base         69s    2.57%        19.68%             10.74%      12.098 K/sec
> this_series  41s    0.62%        12.11%              5.78%       8.617 K/sec
> 
> nr_group=32:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base     192s±25%  15.12%        25.83%             9.33%       12.141 K/sec
> this_series  71s    0.47%        10.98%             4.58%        8.198 K/sec
> 
> I also tested applying my "ratelimit update to tg->load_avg" patch and
> the test summary is:
> - performance improved noticeably for nr_group=20 and slightly for
>    nr_group=10 case; nr_group=32's performance is roughly the same.
> - task migrations dropped for all three cases; nr_group=20 saw the
>    biggest drop.
> - rq lock contention dropped for all three cases and again, nr_group=32
>    saw the biggest drop.
> 
> Below are detailed data.
> Base: peter's sched/core branch with my "ratelimit" patch.
> this_series: apply this patchset on top of base.
> 
> nr_group=10:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base         36s    0.55%        0.46%              1.43%       15.034 K/sec
> this_series  35s    0.56%        0.52%              1.53%       13.751 K/sec
> 
> nr_group=20:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base         47s    1.28%        0.73%              2.33%       21.217 K/sec
> this_series  42s    0.60%        0.69%              1.69%       14.130 K/sec
> 
> nr_group=32:
>              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> base         70s    2.38%        0.60%              2.19%       17.855 K/sec
> this_series  70s    0.58%        0.63%              1.77%       12.331 K/sec
> 
> Thanks,
> Aaron

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
> On 8/24/23 03:52, Aaron Lu wrote:
> > On Wed, Aug 23, 2023 at 02:52:17PM -0400, Mathieu Desnoyers wrote:
> > > On 8/23/23 11:26, Mathieu Desnoyers wrote:
> > > > On 8/22/23 07:31, Mathieu Desnoyers wrote:
> > > > > Introduce cpus_share_l2c to allow querying whether two logical CPUs
> > > > > share a common L2 cache.
> > > > > 
> > > > > Considering a system like the AMD EPYC 9654 96-Core Processor, the L1
> > > > > cache has a latency of 4-5 cycles, the L2 cache has a latency of at
> > > > > least 14ns, whereas the L3 cache has a latency of 50ns [1]. Compared to
> > > > > this, I measured the RAM accesses to a latency around 120ns on my
> > > > > system [2]. So L3 really is only 2.4x faster than RAM accesses.
> > > > > Therefore, with this relatively slow access speed compared to L2, the
> > > > > scheduler will benefit from only considering CPUs sharing an L2 cache
> > > > > for the purpose of using remote runqueue locking rather than queued
> > > > > wakeups.
> > > > 
> > > > So I did some more benchmarking to figure out whether the reason for
> > > > this speedup is the latency delta between L2 and L3, or is due to the
> > > > number of hw threads contending on the rq locks.
> > > > 
> > > > I tried to force grouping of those "skip ttwu queue" groups by a subset
> > > > of the LLC id, basically by taking the LLC id and adding the cpu number
> > > > modulo N, where N is chosen based on my machine topology.
> > > > 
> > > > The end result is that I have similar numbers for groups of 1, 2, 4 HW
> > > > threads (which use rq locks and skip queued ttwu within the group).
> > > > Starting with group of size 8, the performance starts to degrade.
> > > > 
> > > > So I wonder: do machines with more than 4 HW threads per L2 cache exist?
> > > > If it's the case, there we should think about grouping not only by L2
> > > > cache, but also sub-divide this group so the number of hw threads per
> > > > group is at most 4.
> > > > 
> > > > Here are my results with the hackbench test-case:
> > > > 
> > > > Group cpus by 16 hw threads:
> > > > 
> > > > Time: 49s
> > > > 
> > > > - group cpus by 8 hw threads: (llc_id + cpu modulo 2)
> > > > 
> > > > Time: 39s
> > > > 
> > > > - group cpus by 4 hw threads: (llc_id + cpu modulo 4)
> > > > 
> > > > Time: 34s
> > > > 
> > > > - group cpus by 2 hw threads: (llc_id + cpu modulo 8)
> > > > (expect same as L2 grouping on this machine)
> > > > 
> > > > Time: 34s
> > > > 
> > > > - group cpus by 1 hw threads: (cpu)
> > > > 
> > > > Time: 33s
> > > 
> > > One more interesting data point: I tried modifying the grouping
> > > so that I would explicitly group by hw threads which sit in different
> > > L3, and even on different NUMA nodes for some
> > > (group id = cpu_id % 192). This is expected to generate really _bad_
> > > cache locality for the runqueue locks within a group.
> > > 
> > > The result for these groups of 3 HW threads is about 33s with the
> > > hackbench benchmark, which seems to confirm that the cause of the
> > > speedup is reduction of the contention on the rq locks by making the
> > > groups smaller, and therefore reducing the likelihood of contention for the
> > > rq locks, rather than by improving cache locality from L3 to L2.
> > 
> > In addition to reduced rq lock contention, I think another reason this
> > improves performance is because it reduced task migration. Not sure if
> > it is the case on your test system, but on my test machine(Intel SPR),
> > task migration number dropped.
> 
> Yes, it's indeed the case on my system as well. It cuts migrations by half
> (9.2K/sec down to 5.0K/sec).
> 
> > Hackbench on Intel SPR(2sockets/112cores/224threads) test summary:
> > - performance improved for all three cases; the more tasks(groups), the
> >    more performance gain;
> 
> Interesting!
> 
> > - task migrations dropped with this series for nr_group=20 and 32
> >    according to 'perf stat'. migration number didn't drop for nr_group=10
> >    but the two update functions' cost dropped which means fewer access to
> >    tg->load_avg and thus, fewer task migrations. This is contradictory
> >    and I can not explain yet;
> 
> Neither can I.
> 
> > - rq lock contention dropped for all three cases and it dropped the most
> >    under more overloaded case: nr_group=32.
> 
> The fact that you observed rq lock contention dropping is a good sign
> that doing more queued wakeups is a good thing compared to allowing
> non-queued wakeups across cpus sharing a whole LLC.
> 
> At this point I'm not sure if the reduction on rq lock contention is mostly
> due to using queued wakeups rather than grabbing remote rq locks, or by a
> side-effet of doing a queued wakeup rather than immediately doing the
> wakeup, which would open a window where the target rq is still considered
> idle by the various code paths within select_task_rq_fair which don't care
> about rq->ttwu_pending.
> 
> > It's not clear to me why this series can reduce task migrations. I doubt
> > it has something to do with more wakelist style wakeup becasue for this
> > test machine, only a single core with two SMT threads share L2 so more
> > wakeups are through wakelist. In wakelist style wakeup, the target rq's
> > ttwu_pending is set and that will make the target cpu as !idle_cpu();
> > This is faster than grabbing the target rq's lock and then increase
> > target rq's nr_running or set target rq's curr to something else than
> > idle. So wakelist style wakeup can make target cpu appear as non idle
> > faster, but I can't connect this with reduced migration yet, I just feel
> > this might be the reason why task migration reduced.
> 
> Many code paths in select_task_rq_fair don't seem to care about
> rq->ttwu_pending.
> 
> In wake_affine_idle, for sync wakeups, if nr_running is 1 on the waker, we
> choose the waker cpu as target.
> 
> In wake_affine_idle, if none of waker or prev wakee cpus are idle, then it
> uses wake_affine_weight to find out which of the waker/prev wakee cpus are
> targets based on their respective load.
> 
> If wake_affine_idle cannot find an idle waker/prev wakee cpu, and if
> wake_affine_weight finds that the prev wakee cpu had a lower load, then
> wake_affine returns the prev wakee cpu as target. This happens even if the
> prev wakee cpu is not idle.
> 
> This "target" cpu is then passed to select_idle_sibling. It expects the
> available_idle_cpu(target) to check again to see whether the target cpu is
> idle. However, it also uses "sched_idle_cpu(target)" which _only_ considers
> nr_running (not ttwu_pending flag). Likewise for the other similar idleness
> checks below in select_idle_sibling for prev and recent_used_cpu. The same
> happens for the case where a per-cpu kthread
> stacks with the wakee.

sched_idle_cpu() mainly concerns with idle policy tasks and if the rq
does not have any idle policy tasks, it will not return true. Since our
tests do not use idle policy tasks, it should never return true.

> I've tried adding checks for rq->ttwu_pending in those code paths on top of
> my patch and I'm still observing the reduction in number of migrations, so
> it's unclear to me how doing more queued wakeups can reduce migrations the
> way it does.

An interesting puzzle.

> I'm starting to think may want to explore explicitly rate limiting task
> migrations as well.
> 
> For instance, we could do something like this:
> 
> Within a 1ms window, if a task is migrated more than once, the following
> wakeups would consider that the prev runqueue should be considered in
> priority (as if it was completely idle) as long as its load is below a given
> threshold.
> 
> So every 1ms tasks have a chance to be migrated to the idlest runqueues, but
> we would then eliminate those frequent migration patterns which end up being
> bad for cache locality.
> 
> Thoughts ?

Not sure if this is a good idea. I had a feeling it could hurt latency..

Thanks,
Aaron

> > 
> > Below are detailed test data.
> > Base: 6.5-rc1.
> > rq_spin%: The percent of raw_spin_rq_lock_nested() as reported by
> >            perf/events=cycles:pp
> > migration: cpu-migrations reported by "perf stat -a -- sleep 5"
> > 
> > The cmdline used is:
> > hackbench -g $nr_group -f 20 --pipe --threads -l 480000 -s 100
> > 
> > nr_group=10:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base         46s    1.32%        20.06%             10.78%      10.227 K/sec
> > this_series  37s    0.57%        15.08%              7.05%      10.722 K/sec
> > 
> > nr_group=20:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base         69s    2.57%        19.68%             10.74%      12.098 K/sec
> > this_series  41s    0.62%        12.11%              5.78%       8.617 K/sec
> > 
> > nr_group=32:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base     192s±25%  15.12%        25.83%             9.33%       12.141 K/sec
> > this_series  71s    0.47%        10.98%             4.58%        8.198 K/sec
> > 
> > I also tested applying my "ratelimit update to tg->load_avg" patch and
> > the test summary is:
> > - performance improved noticeably for nr_group=20 and slightly for
> >    nr_group=10 case; nr_group=32's performance is roughly the same.
> > - task migrations dropped for all three cases; nr_group=20 saw the
> >    biggest drop.
> > - rq lock contention dropped for all three cases and again, nr_group=32
> >    saw the biggest drop.
> > 
> > Below are detailed data.
> > Base: peter's sched/core branch with my "ratelimit" patch.
> > this_series: apply this patchset on top of base.
> > 
> > nr_group=10:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base         36s    0.55%        0.46%              1.43%       15.034 K/sec
> > this_series  35s    0.56%        0.52%              1.53%       13.751 K/sec
> > 
> > nr_group=20:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base         47s    1.28%        0.73%              2.33%       21.217 K/sec
> > this_series  42s    0.60%        0.69%              1.69%       14.130 K/sec
> > 
> > nr_group=32:
> >              time  rq_spin%  update_cfs_group%  update_load_avg% migration
> > base         70s    2.38%        0.60%              2.19%       17.855 K/sec
> > this_series  70s    0.58%        0.63%              1.77%       12.331 K/sec
> > 
> > Thanks,
> > Aaron

Re: [RFC PATCH v3 0/3] sched: Skip queued wakeups only when L2 is shared

[Swapnil Sapkal]

Hello Mathieu,

On 8/22/2023 5:01 PM, Mathieu Desnoyers wrote:
> This series improves performance of scheduler wakeups on large systems
> by skipping queued wakeups only when CPUs share their L2 cache, rather
> than when they share their LLC.
> 
> The speedup mainly reproduces on workloads which have at least *some*
> idle time (because it significantly increases the number of migrations,
> and thus remote wakeups), *and* it needs to have a sufficient load to
> cause contention on the runqueue locks.
> 
> Feedback is welcome,

I ran some micro-benchmarks as part of testing this series. Here are the
observations:

- Hackbench shows improvement with this patch and Aaron's patch with
   6.5-rc1 kernel as the baseline.

- tbench and netperf shows shows some dip in performance with highly
   overloaded case.

- Other micro-benchmarks shows more or less similar performance with
   these patches.

o System Details

- 4th Generation EPYC System
- 2 x 128C/256T
- NPS1 mode

o Kernels

base:	                                6.5.0-rc1
base + mathieu-queued-wakeup:		6.5.0-rc1 + Mathieu's patches [1]
base + aaron-tg-load-avg: 		6.5.0-rc1 + Aaron's patch [2]
base + queued-wakeup + tg-load-avg:     6.5.0-rc1 + Mathieu's patches [1] + Aaron's patch [2]

[References]

[1] "sched: Skip queued wakeups only when L2 is shared"
     (https://lore.kernel.org/all/20230822113133.643238-1-mathieu.desnoyers@efficios.com/)
[2] "Reduce cost of accessing tg->load_avg"
     (https://lore.kernel.org/lkml/20230823060832.454842-1-aaron.lu@intel.com/)

==================================================================
Test          : hackbench
Units         : Time in seconds
Interpretation: Lower is better
Statistic     : AMean
==================================================================
Test:        6.5.0-rc1 (base)    base + mathieu-queued-wakeup       base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
  1-groups:   22.15 (0.00 pct)      22.46 (-1.39 pct)                    22.35 (-0.90 pct)                   21.20 (4.28 pct)
  2-groups:   22.76 (0.00 pct)      21.78 (4.30 pct)                     22.60 (0.70 pct)                    21.90 (3.77 pct)
  4-groups:   22.12 (0.00 pct)      22.02 (0.45 pct)                     22.22 (-0.45 pct)                   21.94 (0.81 pct)
  8-groups:   24.80 (0.00 pct)      22.36 (9.83 pct)                     22.99 (7.29 pct)                    22.00 (11.29 pct)
16-groups:   31.09 (0.00 pct)      21.56 (30.65 pct)                    22.13 (28.81 pct)                   20.60 (33.74 pct)

==================================================================
Test          : tbench
Units         : Throughput
Interpretation: Higher is better
Statistic     : AMean
==================================================================
Clients: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup           base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
     1    261.49 (0.00 pct)       261.18 (-0.11 pct)                     262.29 (0.30 pct)                   257.80 (-1.41 pct)
     2    514.08 (0.00 pct)       521.30 (1.40 pct)                      517.66 (0.69 pct)                   510.96 (-0.60 pct)
     4    1002.51 (0.00 pct)      988.81 (-1.36 pct)                     995.04 (-0.74 pct)                  987.74 (-1.47 pct)
     8    1978.74 (0.00 pct)      1966.60 (-0.61 pct)                    1991.85 (0.66 pct)                  1941.39 (-1.88 pct)
    16    3864.14 (0.00 pct)      3952.03 (2.27 pct)                     3914.80 (1.31 pct)                  3873.88 (0.25 pct)
    32    7473.19 (0.00 pct)      7602.38 (1.72 pct)                     7585.94 (1.50 pct)                  7423.44 (-0.66 pct)
    64    14335.10 (0.00 pct)     14313.17 (-0.15 pct)                   14474.67 (0.97 pct)                 14030.63 (-2.12 pct)
   128    27275.73 (0.00 pct)     25176.80 (-7.69 pct)                   28066.53 (2.89 pct)                 25045.53 (-8.17 pct)
   256    41688.17 (0.00 pct)     44373.40 (6.44 pct)                    43779.37 (5.01 pct)                 41427.00 (-0.62 pct)
   512    137481.33 (0.00 pct)    136466.67 (-0.73 pct)                  134824.00 (-1.93 pct)               141280.00 (2.76 pct)
  1024    140534.00 (0.00 pct)    141916.33 (0.98 pct)                   137008.33 (-2.50 pct)               126319.33 (-10.11 pct)
  2048    145378.00 (0.00 pct)    145479.33 (0.06 pct)                   138763.67 (-4.54 pct)               124471.00 (-14.38 pct)

  ==================================================================
  Test          : netperf
  Units         : Througput
  Interpretation: Higher is better
  Statistic     : AMean
  ==================================================================
                  6.5.0-rc1 (base)    base + mathieu-queued-wakeup       base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
   1-clients:      59642.88 (0.00 pct)        61647.37 (3.36 pct)         61186.24 (2.58 pct)                 59099.11 (-0.91 pct)
   2-clients:      59349.65 (0.00 pct)        60896.01 (2.60 pct)         60582.49 (2.07 pct)                 62738.47 (5.70 pct)
   4-clients:      59197.37 (0.00 pct)        60457.29 (2.12 pct)         63042.52 (6.49 pct)                 60879.58 (2.84 pct)
   8-clients:      61977.66 (0.00 pct)        60389.92 (-2.56 pct)        62078.15 (0.16 pct)                 60314.65 (-2.68 pct)
  16-clients:      61518.83 (0.00 pct)        61143.51 (-0.61 pct)        60946.08 (-0.93 pct)                59388.78 (-3.46 pct)
  32-clients:      58230.81 (0.00 pct)        58653.20 (0.72 pct)         58594.14 (0.62 pct)                 58188.52 (-0.07 pct)
  64-clients:      58050.92 (0.00 pct)        57834.55 (-0.37 pct)        58183.51 (0.22 pct)                 57565.75 (-0.83 pct)
  128-clients:     54324.55 (0.00 pct)        54385.60 (0.11 pct)         54913.43 (1.08 pct)                 53917.11 (-0.75 pct)
  256-clients:     70155.29 (0.00 pct)        69390.68 (-1.08 pct)        70097.50 (-0.08 pct)                64410.66 (-8.18 pct)
  512-clients:     61511.77 (0.00 pct)        61480.99 (-0.05 pct)        54493.82 (-11.40 pct)               46227.05 (-24.84 pct)

==================================================================
Test          : stream-10
Units         : Bandwidth, MB/s
Interpretation: Higher is better
Statistic     : HMean
==================================================================
Test:      6.5.0-rc1 (base)      base + mathieu-queued-wakeup         base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
  Copy:   353336.76 (0.00 pct)       352956.36 (-0.10 pct)               349583.67 (-1.06 pct)               351152.80 (-0.61 pct)
Scale:   353474.88 (0.00 pct)       354582.35 (0.31 pct)                350543.75 (-0.82 pct)               353275.74 (-0.05 pct)
   Add:   371984.24 (0.00 pct)       372824.87 (0.22 pct)                369173.72 (-0.75 pct)               370483.63 (-0.40 pct)
Triad:   372625.41 (0.00 pct)       278389.62 (-25.28 pct)              369504.06 (-0.83 pct)               369070.11 (-0.95 pct)

==================================================================
Test          : stream-100
Units         : Bandwidth, MB/s
Interpretation: Higher is better
Statistic     : HMean
==================================================================
Test:     6.5.0-rc1 (base)        base + mathieu-queued-wakeup       base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
  Copy:   353476.35 (0.00 pct)       354954.50 (0.41 pct)                354614.56 (0.32 pct)                353512.71 (0.01 pct)
Scale:   353214.73 (0.00 pct)       354884.12 (0.47 pct)                355841.17 (0.74 pct)                353220.53 (0.00 pct)
   Add:   370755.48 (0.00 pct)       372292.72 (0.41 pct)                375307.35 (1.22 pct)                369917.77 (-0.22 pct)
Triad:   370652.02 (0.00 pct)       372732.11 (0.56 pct)                375718.85 (1.36 pct)                369926.26 (-0.19 pct)

==================================================================
Test          : schbench (old)
Units         : 99th percentile latency in us
Interpretation: Lower is better
Statistic     : Median
==================================================================
#workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
   1:      56.00 (0.00 pct)        58.00 (-3.57 pct)                      60.00 (-7.14 pct)                   60.00 (-7.14 pct)
   2:      61.00 (0.00 pct)        56.00 (8.19 pct)                       59.00 (3.27 pct)                    60.00 (1.63 pct)
   4:      64.00 (0.00 pct)        62.00 (3.12 pct)                       66.00 (-3.12 pct)                   64.00 (0.00 pct)
   8:      96.00 (0.00 pct)        78.00 (18.75 pct)                      76.00 (20.83 pct)                   93.00 (3.12 pct)
  16:      98.00 (0.00 pct)        95.00 (3.06 pct)                       98.00 (0.00 pct)                    95.00 (3.06 pct)
  32:     137.00 (0.00 pct)       144.00 (-5.10 pct)                     133.00 (2.91 pct)                   130.00 (5.10 pct)
  64:     206.00 (0.00 pct)       210.00 (-1.94 pct)                     200.00 (2.91 pct)                   217.00 (-5.33 pct)
128:     348.00 (0.00 pct)       347.00 (0.28 pct)                      413.00 (-18.67 pct)                 366.00 (-5.17 pct)
256:     679.00 (0.00 pct)       669.00 (1.47 pct)                      669.00 (1.47 pct)                   675.00 (0.58 pct)
512:     1366.00 (0.00 pct)      1366.00 (0.00 pct)                     1442.00 (-5.56 pct)                 1430.00 (-4.68 pct)


==================================================================
Test          : schbench (new)
Units         : 99th percentile latency in us
Interpretation: Lower is better
Statistic     : Median
==================================================================
Metric: wakeup_lat_summary
#workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
   1:      15.00 (0.00 pct)        15.00 (0.00 pct)                       16.00 (-6.66 pct)                   17.00 (-13.33 pct)
   2:      16.00 (0.00 pct)        16.00 (0.00 pct)                       17.00 (-6.25 pct)                   17.00 (-6.25 pct)
   4:      17.00 (0.00 pct)        17.00 (0.00 pct)                       15.00 (11.76 pct)                   17.00 (0.00 pct)
   8:      11.00 (0.00 pct)        13.00 (-18.18 pct)                     11.00 (0.00 pct)                    11.00 (0.00 pct)
  16:      11.00 (0.00 pct)        11.00 (0.00 pct)                       10.00 (9.09 pct)                     9.00 (18.18 pct)
  32:      11.00 (0.00 pct)        11.00 (0.00 pct)                       11.00 (0.00 pct)                    11.00 (0.00 pct)
  64:      10.00 (0.00 pct)        11.00 (-10.00 pct)                     10.00 (0.00 pct)                    10.00 (0.00 pct)
128:      11.00 (0.00 pct)        12.00 (-9.09 pct)                      12.00 (-9.09 pct)                   11.00 (0.00 pct)
256:     117.00 (0.00 pct)       162.00 (-38.46 pct)                     90.00 (23.07 pct)                  103.00 (11.96 pct)
512:     22496.00 (0.00 pct)     21664.00 (3.69 pct)                    22368.00 (0.56 pct)                 21408.00 (4.83 pct)

Metric: request_lat_summary
#workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
   1:     6872.00 (0.00 pct)      6872.00 (0.00 pct)                     6792.00 (1.16 pct)                  6856.00 (0.23 pct)
   2:     6824.00 (0.00 pct)      6824.00 (0.00 pct)                     6872.00 (-0.70 pct)                 6856.00 (-0.46 pct)
   4:     6824.00 (0.00 pct)      6808.00 (0.23 pct)                     6872.00 (-0.70 pct)                 6824.00 (0.00 pct)
   8:     6824.00 (0.00 pct)      6824.00 (0.00 pct)                     6872.00 (-0.70 pct)                 6824.00 (0.00 pct)
  16:     6824.00 (0.00 pct)      6840.00 (-0.23 pct)                    6872.00 (-0.70 pct)                 6840.00 (-0.23 pct)
  32:     6840.00 (0.00 pct)      6840.00 (0.00 pct)                     6888.00 (-0.70 pct)                 6856.00 (-0.23 pct)
  64:     6840.00 (0.00 pct)      6872.00 (-0.46 pct)                    6888.00 (-0.70 pct)                 6872.00 (-0.46 pct)
128:     12272.00 (0.00 pct)     12784.00 (-4.17 pct)                   13200.00 (-7.56 pct)                12016.00 (2.08 pct)
256:     13328.00 (0.00 pct)     13392.00 (-0.48 pct)                   13712.00 (-2.88 pct)                13552.00 (-1.68 pct)
512:     88832.00 (0.00 pct)     86400.00 (2.73 pct)                    88192.00 (0.72 pct)                 85632.00 (3.60 pct)

Metric: rps_summary
#workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
   1:     297.00 (0.00 pct)       297.00 (0.00 pct)                      297.00 (0.00 pct)                   299.00 (-0.67 pct)
   2:     601.00 (0.00 pct)       603.00 (-0.33 pct)                     595.00 (0.99 pct)                   601.00 (0.00 pct)
   4:     1206.00 (0.00 pct)      1206.00 (0.00 pct)                     1190.00 (1.32 pct)                  1206.00 (0.00 pct)
   8:     2412.00 (0.00 pct)      2412.00 (0.00 pct)                     2396.00 (0.66 pct)                  2420.00 (-0.33 pct)
  16:     4840.00 (0.00 pct)      4824.00 (0.33 pct)                     4792.00 (0.99 pct)                  4840.00 (0.00 pct)
  32:     9648.00 (0.00 pct)      9648.00 (0.00 pct)                     9584.00 (0.66 pct)                  9680.00 (-0.33 pct)
  64:     19360.00 (0.00 pct)     19296.00 (0.33 pct)                    19168.00 (0.99 pct)                 19296.00 (0.33 pct)
128:     37952.00 (0.00 pct)     35264.00 (7.08 pct)                    36672.00 (3.37 pct)                 38080.00 (-0.33 pct)
256:     41408.00 (0.00 pct)     41536.00 (-0.30 pct)                   39744.00 (4.01 pct)                 40896.00 (1.23 pct)
512:     36288.00 (0.00 pct)     36800.00 (-1.41 pct)                   35264.00 (2.82 pct)                 35776.00 (1.41 pct)

Tested-by: Swapnil Sapkal <Swapnil.Sapkal@amd.com>

> 
> Thanks,
> 
> Mathieu
> 
> Mathieu Desnoyers (3):
>    sched: Rename cpus_share_cache to cpus_share_llc
>    sched: Introduce cpus_share_l2c (v3)
>    sched: ttwu_queue_cond: skip queued wakeups across different l2 caches
> 
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Valentin Schneider <vschneid@redhat.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Ben Segall <bsegall@google.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
> Cc: Aaron Lu <aaron.lu@intel.com>
> Cc: Julien Desfossez <jdesfossez@digitalocean.com>
> Cc: x86@kernel.org
> 
>   block/blk-mq.c                 |  2 +-
>   include/linux/sched/topology.h | 10 ++++++++--
>   kernel/sched/core.c            | 14 +++++++++++---
>   kernel/sched/fair.c            |  8 ++++----
>   kernel/sched/sched.h           |  2 ++
>   kernel/sched/topology.c        | 32 +++++++++++++++++++++++++++++---
>   6 files changed, 55 insertions(+), 13 deletions(-)
> 
--
Thanks and Regards,
Swapnil

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

On 8/25/23 02:49, Aaron Lu wrote:
> On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
[...]
>>> - task migrations dropped with this series for nr_group=20 and 32
>>>     according to 'perf stat'. migration number didn't drop for nr_group=10
>>>     but the two update functions' cost dropped which means fewer access to
>>>     tg->load_avg and thus, fewer task migrations. This is contradictory
>>>     and I can not explain yet;
>>
>> Neither can I.
>>

[...]

>>
>>> It's not clear to me why this series can reduce task migrations. I doubt
>>> it has something to do with more wakelist style wakeup becasue for this
>>> test machine, only a single core with two SMT threads share L2 so more
>>> wakeups are through wakelist. In wakelist style wakeup, the target rq's
>>> ttwu_pending is set and that will make the target cpu as !idle_cpu();
>>> This is faster than grabbing the target rq's lock and then increase
>>> target rq's nr_running or set target rq's curr to something else than
>>> idle. So wakelist style wakeup can make target cpu appear as non idle
>>> faster, but I can't connect this with reduced migration yet, I just feel
>>> this might be the reason why task migration reduced.
>>
> 
[...]
>> I've tried adding checks for rq->ttwu_pending in those code paths on top of
>> my patch and I'm still observing the reduction in number of migrations, so
>> it's unclear to me how doing more queued wakeups can reduce migrations the
>> way it does.
> 
> An interesting puzzle.

One metric that can help understand the impact of my patch: comparing
hackbench from a baseline where only your load_avg patch is applied
to a kernel with my l2c patch applied, I notice that the goidle
schedstat is cut in half. For a given CPU (they are pretty much alike),
it goes from 650456 to 353487.

So could it be that by doing queued wakeups, we end up batching
execution of the woken up tasks for a given CPU, rather than going
back and forth between idle and non-idle ? One important thing that
this changes is to reduce the number of newidle balance triggered.

Thoughts ?

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 0/3] sched: Skip queued wakeups only when L2 is shared

[Mathieu Desnoyers]

On 8/25/23 06:11, Swapnil Sapkal wrote:
> Hello Mathieu,
> 
> On 8/22/2023 5:01 PM, Mathieu Desnoyers wrote:
>> This series improves performance of scheduler wakeups on large systems
>> by skipping queued wakeups only when CPUs share their L2 cache, rather
>> than when they share their LLC.
>>
>> The speedup mainly reproduces on workloads which have at least *some*
>> idle time (because it significantly increases the number of migrations,
>> and thus remote wakeups), *and* it needs to have a sufficient load to
>> cause contention on the runqueue locks.
>>
>> Feedback is welcome,
> 
> I ran some micro-benchmarks as part of testing this series. Here are the
> observations:
> 
> - Hackbench shows improvement with this patch and Aaron's patch with
>    6.5-rc1 kernel as the baseline.
> 
> - tbench and netperf shows shows some dip in performance with highly
>    overloaded case.
> 
> - Other micro-benchmarks shows more or less similar performance with
>    these patches.

Those results look promising! Thanks for testing!

Mathieu


> 
> o System Details
> 
> - 4th Generation EPYC System
> - 2 x 128C/256T
> - NPS1 mode
> 
> o Kernels
> 
> base:                                    6.5.0-rc1
> base + mathieu-queued-wakeup:        6.5.0-rc1 + Mathieu's patches [1]
> base + aaron-tg-load-avg:         6.5.0-rc1 + Aaron's patch [2]
> base + queued-wakeup + tg-load-avg:     6.5.0-rc1 + Mathieu's patches 
> [1] + Aaron's patch [2]
> 
> [References]
> 
> [1] "sched: Skip queued wakeups only when L2 is shared"
>      
> (https://lore.kernel.org/all/20230822113133.643238-1-mathieu.desnoyers@efficios.com/)
> [2] "Reduce cost of accessing tg->load_avg"
>      
> (https://lore.kernel.org/lkml/20230823060832.454842-1-aaron.lu@intel.com/)
> 
> ==================================================================
> Test          : hackbench
> Units         : Time in seconds
> Interpretation: Lower is better
> Statistic     : AMean
> ==================================================================
> Test:        6.5.0-rc1 (base)    base + mathieu-queued-wakeup       base 
> + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
>   1-groups:   22.15 (0.00 pct)      22.46 (-1.39 pct)                    
> 22.35 (-0.90 pct)                   21.20 (4.28 pct)
>   2-groups:   22.76 (0.00 pct)      21.78 (4.30 pct)                     
> 22.60 (0.70 pct)                    21.90 (3.77 pct)
>   4-groups:   22.12 (0.00 pct)      22.02 (0.45 pct)                     
> 22.22 (-0.45 pct)                   21.94 (0.81 pct)
>   8-groups:   24.80 (0.00 pct)      22.36 (9.83 pct)                     
> 22.99 (7.29 pct)                    22.00 (11.29 pct)
> 16-groups:   31.09 (0.00 pct)      21.56 (30.65 pct)                    
> 22.13 (28.81 pct)                   20.60 (33.74 pct)
> 
> ==================================================================
> Test          : tbench
> Units         : Throughput
> Interpretation: Higher is better
> Statistic     : AMean
> ==================================================================
> Clients: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup           base 
> + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
>      1    261.49 (0.00 pct)       261.18 (-0.11 pct)                     
> 262.29 (0.30 pct)                   257.80 (-1.41 pct)
>      2    514.08 (0.00 pct)       521.30 (1.40 pct)                      
> 517.66 (0.69 pct)                   510.96 (-0.60 pct)
>      4    1002.51 (0.00 pct)      988.81 (-1.36 pct)                     
> 995.04 (-0.74 pct)                  987.74 (-1.47 pct)
>      8    1978.74 (0.00 pct)      1966.60 (-0.61 pct)                    
> 1991.85 (0.66 pct)                  1941.39 (-1.88 pct)
>     16    3864.14 (0.00 pct)      3952.03 (2.27 pct)                     
> 3914.80 (1.31 pct)                  3873.88 (0.25 pct)
>     32    7473.19 (0.00 pct)      7602.38 (1.72 pct)                     
> 7585.94 (1.50 pct)                  7423.44 (-0.66 pct)
>     64    14335.10 (0.00 pct)     14313.17 (-0.15 pct)                   
> 14474.67 (0.97 pct)                 14030.63 (-2.12 pct)
>    128    27275.73 (0.00 pct)     25176.80 (-7.69 pct)                   
> 28066.53 (2.89 pct)                 25045.53 (-8.17 pct)
>    256    41688.17 (0.00 pct)     44373.40 (6.44 pct)                    
> 43779.37 (5.01 pct)                 41427.00 (-0.62 pct)
>    512    137481.33 (0.00 pct)    136466.67 (-0.73 pct)                  
> 134824.00 (-1.93 pct)               141280.00 (2.76 pct)
>   1024    140534.00 (0.00 pct)    141916.33 (0.98 pct)                   
> 137008.33 (-2.50 pct)               126319.33 (-10.11 pct)
>   2048    145378.00 (0.00 pct)    145479.33 (0.06 pct)                   
> 138763.67 (-4.54 pct)               124471.00 (-14.38 pct)
> 
>   ==================================================================
>   Test          : netperf
>   Units         : Througput
>   Interpretation: Higher is better
>   Statistic     : AMean
>   ==================================================================
>                   6.5.0-rc1 (base)    base + mathieu-queued-wakeup       
> base + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
>    1-clients:      59642.88 (0.00 pct)        61647.37 (3.36 
> pct)         61186.24 (2.58 pct)                 59099.11 (-0.91 pct)
>    2-clients:      59349.65 (0.00 pct)        60896.01 (2.60 
> pct)         60582.49 (2.07 pct)                 62738.47 (5.70 pct)
>    4-clients:      59197.37 (0.00 pct)        60457.29 (2.12 
> pct)         63042.52 (6.49 pct)                 60879.58 (2.84 pct)
>    8-clients:      61977.66 (0.00 pct)        60389.92 (-2.56 
> pct)        62078.15 (0.16 pct)                 60314.65 (-2.68 pct)
>   16-clients:      61518.83 (0.00 pct)        61143.51 (-0.61 
> pct)        60946.08 (-0.93 pct)                59388.78 (-3.46 pct)
>   32-clients:      58230.81 (0.00 pct)        58653.20 (0.72 
> pct)         58594.14 (0.62 pct)                 58188.52 (-0.07 pct)
>   64-clients:      58050.92 (0.00 pct)        57834.55 (-0.37 
> pct)        58183.51 (0.22 pct)                 57565.75 (-0.83 pct)
>   128-clients:     54324.55 (0.00 pct)        54385.60 (0.11 
> pct)         54913.43 (1.08 pct)                 53917.11 (-0.75 pct)
>   256-clients:     70155.29 (0.00 pct)        69390.68 (-1.08 
> pct)        70097.50 (-0.08 pct)                64410.66 (-8.18 pct)
>   512-clients:     61511.77 (0.00 pct)        61480.99 (-0.05 
> pct)        54493.82 (-11.40 pct)               46227.05 (-24.84 pct)
> 
> ==================================================================
> Test          : stream-10
> Units         : Bandwidth, MB/s
> Interpretation: Higher is better
> Statistic     : HMean
> ==================================================================
> Test:      6.5.0-rc1 (base)      base + mathieu-queued-wakeup         
> base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
>   Copy:   353336.76 (0.00 pct)       352956.36 (-0.10 pct)               
> 349583.67 (-1.06 pct)               351152.80 (-0.61 pct)
> Scale:   353474.88 (0.00 pct)       354582.35 (0.31 pct)                
> 350543.75 (-0.82 pct)               353275.74 (-0.05 pct)
>    Add:   371984.24 (0.00 pct)       372824.87 (0.22 pct)                
> 369173.72 (-0.75 pct)               370483.63 (-0.40 pct)
> Triad:   372625.41 (0.00 pct)       278389.62 (-25.28 pct)              
> 369504.06 (-0.83 pct)               369070.11 (-0.95 pct)
> 
> ==================================================================
> Test          : stream-100
> Units         : Bandwidth, MB/s
> Interpretation: Higher is better
> Statistic     : HMean
> ==================================================================
> Test:     6.5.0-rc1 (base)        base + mathieu-queued-wakeup       
> base + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
>   Copy:   353476.35 (0.00 pct)       354954.50 (0.41 pct)                
> 354614.56 (0.32 pct)                353512.71 (0.01 pct)
> Scale:   353214.73 (0.00 pct)       354884.12 (0.47 pct)                
> 355841.17 (0.74 pct)                353220.53 (0.00 pct)
>    Add:   370755.48 (0.00 pct)       372292.72 (0.41 pct)                
> 375307.35 (1.22 pct)                369917.77 (-0.22 pct)
> Triad:   370652.02 (0.00 pct)       372732.11 (0.56 pct)                
> 375718.85 (1.36 pct)                369926.26 (-0.19 pct)
> 
> ==================================================================
> Test          : schbench (old)
> Units         : 99th percentile latency in us
> Interpretation: Lower is better
> Statistic     : Median
> ==================================================================
> #workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base 
> + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
>    1:      56.00 (0.00 pct)        58.00 (-3.57 
> pct)                      60.00 (-7.14 pct)                   60.00 
> (-7.14 pct)
>    2:      61.00 (0.00 pct)        56.00 (8.19 
> pct)                       59.00 (3.27 pct)                    60.00 
> (1.63 pct)
>    4:      64.00 (0.00 pct)        62.00 (3.12 
> pct)                       66.00 (-3.12 pct)                   64.00 
> (0.00 pct)
>    8:      96.00 (0.00 pct)        78.00 (18.75 
> pct)                      76.00 (20.83 pct)                   93.00 
> (3.12 pct)
>   16:      98.00 (0.00 pct)        95.00 (3.06 
> pct)                       98.00 (0.00 pct)                    95.00 
> (3.06 pct)
>   32:     137.00 (0.00 pct)       144.00 (-5.10 pct)                     
> 133.00 (2.91 pct)                   130.00 (5.10 pct)
>   64:     206.00 (0.00 pct)       210.00 (-1.94 pct)                     
> 200.00 (2.91 pct)                   217.00 (-5.33 pct)
> 128:     348.00 (0.00 pct)       347.00 (0.28 pct)                      
> 413.00 (-18.67 pct)                 366.00 (-5.17 pct)
> 256:     679.00 (0.00 pct)       669.00 (1.47 pct)                      
> 669.00 (1.47 pct)                   675.00 (0.58 pct)
> 512:     1366.00 (0.00 pct)      1366.00 (0.00 pct)                     
> 1442.00 (-5.56 pct)                 1430.00 (-4.68 pct)
> 
> 
> ==================================================================
> Test          : schbench (new)
> Units         : 99th percentile latency in us
> Interpretation: Lower is better
> Statistic     : Median
> ==================================================================
> Metric: wakeup_lat_summary
> #workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base 
> + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
>    1:      15.00 (0.00 pct)        15.00 (0.00 
> pct)                       16.00 (-6.66 pct)                   17.00 
> (-13.33 pct)
>    2:      16.00 (0.00 pct)        16.00 (0.00 
> pct)                       17.00 (-6.25 pct)                   17.00 
> (-6.25 pct)
>    4:      17.00 (0.00 pct)        17.00 (0.00 
> pct)                       15.00 (11.76 pct)                   17.00 
> (0.00 pct)
>    8:      11.00 (0.00 pct)        13.00 (-18.18 
> pct)                     11.00 (0.00 pct)                    11.00 (0.00 
> pct)
>   16:      11.00 (0.00 pct)        11.00 (0.00 
> pct)                       10.00 (9.09 pct)                     9.00 
> (18.18 pct)
>   32:      11.00 (0.00 pct)        11.00 (0.00 
> pct)                       11.00 (0.00 pct)                    11.00 
> (0.00 pct)
>   64:      10.00 (0.00 pct)        11.00 (-10.00 
> pct)                     10.00 (0.00 pct)                    10.00 (0.00 
> pct)
> 128:      11.00 (0.00 pct)        12.00 (-9.09 pct)                      
> 12.00 (-9.09 pct)                   11.00 (0.00 pct)
> 256:     117.00 (0.00 pct)       162.00 (-38.46 pct)                     
> 90.00 (23.07 pct)                  103.00 (11.96 pct)
> 512:     22496.00 (0.00 pct)     21664.00 (3.69 pct)                    
> 22368.00 (0.56 pct)                 21408.00 (4.83 pct)
> 
> Metric: request_lat_summary
> #workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base 
> + aaron-tg-load-avg        base + queued-wakeup + tg-load-avg
>    1:     6872.00 (0.00 pct)      6872.00 (0.00 pct)                     
> 6792.00 (1.16 pct)                  6856.00 (0.23 pct)
>    2:     6824.00 (0.00 pct)      6824.00 (0.00 pct)                     
> 6872.00 (-0.70 pct)                 6856.00 (-0.46 pct)
>    4:     6824.00 (0.00 pct)      6808.00 (0.23 pct)                     
> 6872.00 (-0.70 pct)                 6824.00 (0.00 pct)
>    8:     6824.00 (0.00 pct)      6824.00 (0.00 pct)                     
> 6872.00 (-0.70 pct)                 6824.00 (0.00 pct)
>   16:     6824.00 (0.00 pct)      6840.00 (-0.23 pct)                    
> 6872.00 (-0.70 pct)                 6840.00 (-0.23 pct)
>   32:     6840.00 (0.00 pct)      6840.00 (0.00 pct)                     
> 6888.00 (-0.70 pct)                 6856.00 (-0.23 pct)
>   64:     6840.00 (0.00 pct)      6872.00 (-0.46 pct)                    
> 6888.00 (-0.70 pct)                 6872.00 (-0.46 pct)
> 128:     12272.00 (0.00 pct)     12784.00 (-4.17 pct)                   
> 13200.00 (-7.56 pct)                12016.00 (2.08 pct)
> 256:     13328.00 (0.00 pct)     13392.00 (-0.48 pct)                   
> 13712.00 (-2.88 pct)                13552.00 (-1.68 pct)
> 512:     88832.00 (0.00 pct)     86400.00 (2.73 pct)                    
> 88192.00 (0.72 pct)                 85632.00 (3.60 pct)
> 
> Metric: rps_summary
> #workers: 6.5.0-rc1 (base)    base + mathieu-queued-wakeup          base 
> + aaron-tg-load-avg       base + queued-wakeup + tg-load-avg
>    1:     297.00 (0.00 pct)       297.00 (0.00 pct)                      
> 297.00 (0.00 pct)                   299.00 (-0.67 pct)
>    2:     601.00 (0.00 pct)       603.00 (-0.33 pct)                     
> 595.00 (0.99 pct)                   601.00 (0.00 pct)
>    4:     1206.00 (0.00 pct)      1206.00 (0.00 pct)                     
> 1190.00 (1.32 pct)                  1206.00 (0.00 pct)
>    8:     2412.00 (0.00 pct)      2412.00 (0.00 pct)                     
> 2396.00 (0.66 pct)                  2420.00 (-0.33 pct)
>   16:     4840.00 (0.00 pct)      4824.00 (0.33 pct)                     
> 4792.00 (0.99 pct)                  4840.00 (0.00 pct)
>   32:     9648.00 (0.00 pct)      9648.00 (0.00 pct)                     
> 9584.00 (0.66 pct)                  9680.00 (-0.33 pct)
>   64:     19360.00 (0.00 pct)     19296.00 (0.33 pct)                    
> 19168.00 (0.99 pct)                 19296.00 (0.33 pct)
> 128:     37952.00 (0.00 pct)     35264.00 (7.08 pct)                    
> 36672.00 (3.37 pct)                 38080.00 (-0.33 pct)
> 256:     41408.00 (0.00 pct)     41536.00 (-0.30 pct)                   
> 39744.00 (4.01 pct)                 40896.00 (1.23 pct)
> 512:     36288.00 (0.00 pct)     36800.00 (-1.41 pct)                   
> 35264.00 (2.82 pct)                 35776.00 (1.41 pct)
> 
> Tested-by: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
> 
>>
>> Thanks,
>>
>> Mathieu
>>
>> Mathieu Desnoyers (3):
>>    sched: Rename cpus_share_cache to cpus_share_llc
>>    sched: Introduce cpus_share_l2c (v3)
>>    sched: ttwu_queue_cond: skip queued wakeups across different l2 caches
>>
>> Cc: Ingo Molnar <mingo@redhat.com>
>> Cc: Peter Zijlstra <peterz@infradead.org>
>> Cc: Valentin Schneider <vschneid@redhat.com>
>> Cc: Steven Rostedt <rostedt@goodmis.org>
>> Cc: Ben Segall <bsegall@google.com>
>> Cc: Mel Gorman <mgorman@suse.de>
>> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
>> Cc: Vincent Guittot <vincent.guittot@linaro.org>
>> Cc: Juri Lelli <juri.lelli@redhat.com>
>> Cc: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
>> Cc: Aaron Lu <aaron.lu@intel.com>
>> Cc: Julien Desfossez <jdesfossez@digitalocean.com>
>> Cc: x86@kernel.org
>>
>>   block/blk-mq.c                 |  2 +-
>>   include/linux/sched/topology.h | 10 ++++++++--
>>   kernel/sched/core.c            | 14 +++++++++++---
>>   kernel/sched/fair.c            |  8 ++++----
>>   kernel/sched/sched.h           |  2 ++
>>   kernel/sched/topology.c        | 32 +++++++++++++++++++++++++++++---
>>   6 files changed, 55 insertions(+), 13 deletions(-)
>>
> -- 
> Thanks and Regards,
> Swapnil

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Fri, Aug 25, 2023 at 09:51:19AM -0400, Mathieu Desnoyers wrote:
> On 8/25/23 02:49, Aaron Lu wrote:
> > On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
> [...]
> > > > - task migrations dropped with this series for nr_group=20 and 32
> > > >     according to 'perf stat'. migration number didn't drop for nr_group=10
> > > >     but the two update functions' cost dropped which means fewer access to
> > > >     tg->load_avg and thus, fewer task migrations. This is contradictory
> > > >     and I can not explain yet;
> > > 
> > > Neither can I.
> > > 
> 
> [...]
> 
> > > 
> > > > It's not clear to me why this series can reduce task migrations. I doubt
> > > > it has something to do with more wakelist style wakeup becasue for this
> > > > test machine, only a single core with two SMT threads share L2 so more
> > > > wakeups are through wakelist. In wakelist style wakeup, the target rq's
> > > > ttwu_pending is set and that will make the target cpu as !idle_cpu();
> > > > This is faster than grabbing the target rq's lock and then increase
> > > > target rq's nr_running or set target rq's curr to something else than
> > > > idle. So wakelist style wakeup can make target cpu appear as non idle
> > > > faster, but I can't connect this with reduced migration yet, I just feel
> > > > this might be the reason why task migration reduced.
> > > 
> > 
> [...]
> > > I've tried adding checks for rq->ttwu_pending in those code paths on top of
> > > my patch and I'm still observing the reduction in number of migrations, so
> > > it's unclear to me how doing more queued wakeups can reduce migrations the
> > > way it does.
> > 
> > An interesting puzzle.
> 
> One metric that can help understand the impact of my patch: comparing
> hackbench from a baseline where only your load_avg patch is applied
> to a kernel with my l2c patch applied, I notice that the goidle
> schedstat is cut in half. For a given CPU (they are pretty much alike),
> it goes from 650456 to 353487.
> 
> So could it be that by doing queued wakeups, we end up batching
> execution of the woken up tasks for a given CPU, rather than going
> back and forth between idle and non-idle ? One important thing that
> this changes is to reduce the number of newidle balance triggered.

I noticed the majority(>99%) migrations are from wakeup path on this
Intel SPR when running hackbench: ttwu() -> set_task_cpu() ->
migrate_task_rq_fair(), so while I think it's a good finding that
newidle balance dropped, it's probably not the reason why migration
number dropped...

Thanks,
Aaron

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Mon, Aug 28, 2023 at 07:19:45PM +0800, Aaron Lu wrote:
> On Fri, Aug 25, 2023 at 09:51:19AM -0400, Mathieu Desnoyers wrote:
> > On 8/25/23 02:49, Aaron Lu wrote:
> > > On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
> > [...]
> > > > > - task migrations dropped with this series for nr_group=20 and 32
> > > > >     according to 'perf stat'. migration number didn't drop for nr_group=10
> > > > >     but the two update functions' cost dropped which means fewer access to
> > > > >     tg->load_avg and thus, fewer task migrations. This is contradictory
> > > > >     and I can not explain yet;
> > > > 
> > > > Neither can I.
> > > > 
> > 
> > [...]
> > 
> > > > 
> > > > > It's not clear to me why this series can reduce task migrations. I doubt
> > > > > it has something to do with more wakelist style wakeup becasue for this
> > > > > test machine, only a single core with two SMT threads share L2 so more
> > > > > wakeups are through wakelist. In wakelist style wakeup, the target rq's
> > > > > ttwu_pending is set and that will make the target cpu as !idle_cpu();
> > > > > This is faster than grabbing the target rq's lock and then increase
> > > > > target rq's nr_running or set target rq's curr to something else than
> > > > > idle. So wakelist style wakeup can make target cpu appear as non idle
> > > > > faster, but I can't connect this with reduced migration yet, I just feel
> > > > > this might be the reason why task migration reduced.
> > > > 
> > > 
> > [...]
> > > > I've tried adding checks for rq->ttwu_pending in those code paths on top of
> > > > my patch and I'm still observing the reduction in number of migrations, so
> > > > it's unclear to me how doing more queued wakeups can reduce migrations the
> > > > way it does.
> > > 
> > > An interesting puzzle.
> > 
> > One metric that can help understand the impact of my patch: comparing
> > hackbench from a baseline where only your load_avg patch is applied
> > to a kernel with my l2c patch applied, I notice that the goidle
> > schedstat is cut in half. For a given CPU (they are pretty much alike),
> > it goes from 650456 to 353487.
> > 
> > So could it be that by doing queued wakeups, we end up batching
> > execution of the woken up tasks for a given CPU, rather than going
> > back and forth between idle and non-idle ? One important thing that
> > this changes is to reduce the number of newidle balance triggered.
> 
> I noticed the majority(>99%) migrations are from wakeup path on this
> Intel SPR when running hackbench: ttwu() -> set_task_cpu() ->
> migrate_task_rq_fair(), so while I think it's a good finding that
> newidle balance dropped, it's probably not the reason why migration
> number dropped...

I profiled select_idle_sibling() and found that with this series,
select_idle_cpu() tends to fail more and select_idle_sibling() fallbacks
to use target in the end, which equals to prev_cpu very often.

Initially I think the reason why select_idle_cpu() failed more with this
series is because "wake_list style enqueue" can make the target cpu appear
as busy earlier and thus, it will be harder for select_idle_cpu() to
find an idle cpu overall. But I also suspect SIS_UTIL makes a difference
here: in vanilla kernel, the idle% is 8% and with this series, the idle%
is only 2% and SIS_UTIL may simply skip doing any search for idle cpu.
Anyway, I think I'll also need to profile select_idle_cpu() to see
what's going on there too.

The above profile was done with below workload on a 2 sockets Intel SPR:
hackbench -g 20 -f 20 --pipe --threads -l 480000 -s 100

Thanks,
Aaron

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Fri, Sep 01, 2023 at 09:45:28PM +0800, Aaron Lu wrote:
> On Mon, Aug 28, 2023 at 07:19:45PM +0800, Aaron Lu wrote:
> > On Fri, Aug 25, 2023 at 09:51:19AM -0400, Mathieu Desnoyers wrote:
> > > On 8/25/23 02:49, Aaron Lu wrote:
> > > > On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
> > > [...]
> > > > > > - task migrations dropped with this series for nr_group=20 and 32
> > > > > >     according to 'perf stat'. migration number didn't drop for nr_group=10
> > > > > >     but the two update functions' cost dropped which means fewer access to
> > > > > >     tg->load_avg and thus, fewer task migrations. This is contradictory
> > > > > >     and I can not explain yet;
> > > > > 
> > > > > Neither can I.
> > > > > 
> > > 
> > > [...]
> > > 
> > > > > 
> > > > > > It's not clear to me why this series can reduce task migrations. I doubt
> > > > > > it has something to do with more wakelist style wakeup becasue for this
> > > > > > test machine, only a single core with two SMT threads share L2 so more
> > > > > > wakeups are through wakelist. In wakelist style wakeup, the target rq's
> > > > > > ttwu_pending is set and that will make the target cpu as !idle_cpu();
> > > > > > This is faster than grabbing the target rq's lock and then increase
> > > > > > target rq's nr_running or set target rq's curr to something else than
> > > > > > idle. So wakelist style wakeup can make target cpu appear as non idle
> > > > > > faster, but I can't connect this with reduced migration yet, I just feel
> > > > > > this might be the reason why task migration reduced.
> > > > > 
> > > > 
> > > [...]
> > > > > I've tried adding checks for rq->ttwu_pending in those code paths on top of
> > > > > my patch and I'm still observing the reduction in number of migrations, so
> > > > > it's unclear to me how doing more queued wakeups can reduce migrations the
> > > > > way it does.
> > > > 
> > > > An interesting puzzle.
> > > 
> > > One metric that can help understand the impact of my patch: comparing
> > > hackbench from a baseline where only your load_avg patch is applied
> > > to a kernel with my l2c patch applied, I notice that the goidle
> > > schedstat is cut in half. For a given CPU (they are pretty much alike),
> > > it goes from 650456 to 353487.
> > > 
> > > So could it be that by doing queued wakeups, we end up batching
> > > execution of the woken up tasks for a given CPU, rather than going
> > > back and forth between idle and non-idle ? One important thing that
> > > this changes is to reduce the number of newidle balance triggered.
> > 
> > I noticed the majority(>99%) migrations are from wakeup path on this
> > Intel SPR when running hackbench: ttwu() -> set_task_cpu() ->
> > migrate_task_rq_fair(), so while I think it's a good finding that
> > newidle balance dropped, it's probably not the reason why migration
> > number dropped...
> 
> I profiled select_idle_sibling() and found that with this series,
> select_idle_cpu() tends to fail more and select_idle_sibling() fallbacks
> to use target in the end, which equals to prev_cpu very often.
> 
> Initially I think the reason why select_idle_cpu() failed more with this
> series is because "wake_list style enqueue" can make the target cpu appear
> as busy earlier and thus, it will be harder for select_idle_cpu() to
> find an idle cpu overall. But I also suspect SIS_UTIL makes a difference
> here: in vanilla kernel, the idle% is 8% and with this series, the idle%
> is only 2% and SIS_UTIL may simply skip doing any search for idle cpu.
> Anyway, I think I'll also need to profile select_idle_cpu() to see
> what's going on there too.

Looks like the reduction in task migration is due to SIS_UTIL, i.e.
select_idle_cpu() aborts a lot more after applying this series because
system utilization increased.

Here are some numbers:
                 @sis       @sic     @migrate_idle_cpu  @abort
vanilla:       24640640   15883958     11913588          4148649
this_series:   22345434   18597564      4294995         14319284

note:
- @sis: number of times select_idle_sibling() called;
- @sic: number of times select_idle_cpu() called;
- @migrate_idle_cpu: number of times task migrated due to
  select_idle_cpu() found an idle cpu that is different from prev_cpu;
- @abort: number of times select_idle_cpu() aborts the search due to
  SIS_UTIL.

All numbers are captured during a 5s window while running the below
workload on a 2 sockets Intel SPR(56 cores, 112 threads per socket):
hackbench -g 20 -f 20 --pipe --threads -l 480000 -s 100

So for this workload, I think this series is doing something good: it
increased system utilization and due to SIS_UTIL, it also reduced task
migration where task migration isn't very useful since system is already
overloaded.

Thanks,
Aaron

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Mathieu Desnoyers]

On 9/5/23 03:21, Aaron Lu wrote:
> On Fri, Sep 01, 2023 at 09:45:28PM +0800, Aaron Lu wrote:
>> On Mon, Aug 28, 2023 at 07:19:45PM +0800, Aaron Lu wrote:
>>> On Fri, Aug 25, 2023 at 09:51:19AM -0400, Mathieu Desnoyers wrote:
>>>> On 8/25/23 02:49, Aaron Lu wrote:
>>>>> On Thu, Aug 24, 2023 at 10:40:45AM -0400, Mathieu Desnoyers wrote:
>>>> [...]
>>>>>>> - task migrations dropped with this series for nr_group=20 and 32
>>>>>>>      according to 'perf stat'. migration number didn't drop for nr_group=10
>>>>>>>      but the two update functions' cost dropped which means fewer access to
>>>>>>>      tg->load_avg and thus, fewer task migrations. This is contradictory
>>>>>>>      and I can not explain yet;
>>>>>>
>>>>>> Neither can I.
>>>>>>
>>>>
>>>> [...]
>>>>
>>>>>>
>>>>>>> It's not clear to me why this series can reduce task migrations. I doubt
>>>>>>> it has something to do with more wakelist style wakeup becasue for this
>>>>>>> test machine, only a single core with two SMT threads share L2 so more
>>>>>>> wakeups are through wakelist. In wakelist style wakeup, the target rq's
>>>>>>> ttwu_pending is set and that will make the target cpu as !idle_cpu();
>>>>>>> This is faster than grabbing the target rq's lock and then increase
>>>>>>> target rq's nr_running or set target rq's curr to something else than
>>>>>>> idle. So wakelist style wakeup can make target cpu appear as non idle
>>>>>>> faster, but I can't connect this with reduced migration yet, I just feel
>>>>>>> this might be the reason why task migration reduced.
>>>>>>
>>>>>
>>>> [...]
>>>>>> I've tried adding checks for rq->ttwu_pending in those code paths on top of
>>>>>> my patch and I'm still observing the reduction in number of migrations, so
>>>>>> it's unclear to me how doing more queued wakeups can reduce migrations the
>>>>>> way it does.
>>>>>
>>>>> An interesting puzzle.
>>>>
>>>> One metric that can help understand the impact of my patch: comparing
>>>> hackbench from a baseline where only your load_avg patch is applied
>>>> to a kernel with my l2c patch applied, I notice that the goidle
>>>> schedstat is cut in half. For a given CPU (they are pretty much alike),
>>>> it goes from 650456 to 353487.
>>>>
>>>> So could it be that by doing queued wakeups, we end up batching
>>>> execution of the woken up tasks for a given CPU, rather than going
>>>> back and forth between idle and non-idle ? One important thing that
>>>> this changes is to reduce the number of newidle balance triggered.
>>>
>>> I noticed the majority(>99%) migrations are from wakeup path on this
>>> Intel SPR when running hackbench: ttwu() -> set_task_cpu() ->
>>> migrate_task_rq_fair(), so while I think it's a good finding that
>>> newidle balance dropped, it's probably not the reason why migration
>>> number dropped...
>>
>> I profiled select_idle_sibling() and found that with this series,
>> select_idle_cpu() tends to fail more and select_idle_sibling() fallbacks
>> to use target in the end, which equals to prev_cpu very often.
>>
>> Initially I think the reason why select_idle_cpu() failed more with this
>> series is because "wake_list style enqueue" can make the target cpu appear
>> as busy earlier and thus, it will be harder for select_idle_cpu() to
>> find an idle cpu overall. But I also suspect SIS_UTIL makes a difference
>> here: in vanilla kernel, the idle% is 8% and with this series, the idle%
>> is only 2% and SIS_UTIL may simply skip doing any search for idle cpu.
>> Anyway, I think I'll also need to profile select_idle_cpu() to see
>> what's going on there too.
> 
> Looks like the reduction in task migration is due to SIS_UTIL, i.e.
> select_idle_cpu() aborts a lot more after applying this series because
> system utilization increased.
> 
> Here are some numbers:
>                   @sis       @sic     @migrate_idle_cpu  @abort
> vanilla:       24640640   15883958     11913588          4148649
> this_series:   22345434   18597564      4294995         14319284
> 
> note:
> - @sis: number of times select_idle_sibling() called;
> - @sic: number of times select_idle_cpu() called;
> - @migrate_idle_cpu: number of times task migrated due to
>    select_idle_cpu() found an idle cpu that is different from prev_cpu;
> - @abort: number of times select_idle_cpu() aborts the search due to
>    SIS_UTIL.
> 
> All numbers are captured during a 5s window while running the below
> workload on a 2 sockets Intel SPR(56 cores, 112 threads per socket):
> hackbench -g 20 -f 20 --pipe --threads -l 480000 -s 100
> 
> So for this workload, I think this series is doing something good: it
> increased system utilization and due to SIS_UTIL, it also reduced task
> migration where task migration isn't very useful since system is already
> overloaded.

This is interesting. Did you also profile the impact of the patches on 
wake_affine(), especially wake_affine_idle() ? Its behavior did change 
very significantly in my tests, and this impacts the target cpu number 
received by select_idle_sibling(). But independently of what 
wake_affine() returns as target (waker cpu or prev_cpu), if 
select_idle_cpu() is trigger-happy and finds idle cores near that 
target, this will cause lots of migrations.

Based on your metrics, the ttwu-queued-l2 approach (in addition to 
reduce lock contention) appear to decrease the SIS_UTIL idleless level 
of the cpus enough to completely change the runqueue selection and 
migration behavior.

I fear that we hide a bad scheduler behavior under the rug by changing 
the idleless level of a specific workload pattern, while leaving the 
underlying root cause unfixed.

I'm currently working on a different approach: rate limit migrations. 
Basically, the idea is to detect when a task is migrated too often for 
its own good, and prevent the scheduler from migrating it for a short 
while. I get about 30% performance improvement with this approach as 
well (limit migration to 1 per 2ms window per task). I'll finish 
polishing my commit messages and send a series as RFC soon.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Re: [RFC PATCH v3 2/3] sched: Introduce cpus_share_l2c

[Aaron Lu]

On Tue, Sep 05, 2023 at 08:46:42AM -0400, Mathieu Desnoyers wrote:
> On 9/5/23 03:21, Aaron Lu wrote:
> > Looks like the reduction in task migration is due to SIS_UTIL, i.e.
> > select_idle_cpu() aborts a lot more after applying this series because
> > system utilization increased.
> > 
> > Here are some numbers:
> >                   @sis       @sic     @migrate_idle_cpu  @abort
> > vanilla:       24640640   15883958     11913588          4148649
> > this_series:   22345434   18597564      4294995         14319284
> > 
> > note:
> > - @sis: number of times select_idle_sibling() called;
> > - @sic: number of times select_idle_cpu() called;
> > - @migrate_idle_cpu: number of times task migrated due to
> >    select_idle_cpu() found an idle cpu that is different from prev_cpu;
> > - @abort: number of times select_idle_cpu() aborts the search due to
> >    SIS_UTIL.
> > 
> > All numbers are captured during a 5s window while running the below
> > workload on a 2 sockets Intel SPR(56 cores, 112 threads per socket):
> > hackbench -g 20 -f 20 --pipe --threads -l 480000 -s 100
> > 
> > So for this workload, I think this series is doing something good: it
> > increased system utilization and due to SIS_UTIL, it also reduced task
> > migration where task migration isn't very useful since system is already
> > overloaded.
> 
> This is interesting. Did you also profile the impact of the patches on
> wake_affine(), especially wake_affine_idle() ? Its behavior did change very

For group=20 case, wake_affine() and wake_affine_idle() don't appear to
change much on this Intel machine, in that target received by sis() is
mostly prev_cpu instead of waker(this) cpu for both kernels.

But I do notice for group=32 case, in vanilla kernel, the chance of target
as received by sis() becoming to waker cpu increased a lot while with
this series, targer remains mostly prev_cpu and that is the reason why
migration dropped with this series for group=32 case becasue when sis()
fallback to use target, this series has a higher chance of not mirgating
the task. And my profile shows for vanilla kernel, when it choose target
as waker cpu, it's mostly due to wake_affine_weight(), not wake_affine_idle().

Thanks,
Aaron

> significantly in my tests, and this impacts the target cpu number received
> by select_idle_sibling(). But independently of what wake_affine() returns as
> target (waker cpu or prev_cpu), if select_idle_cpu() is trigger-happy and
> finds idle cores near that target, this will cause lots of migrations.
> 
> Based on your metrics, the ttwu-queued-l2 approach (in addition to reduce
> lock contention) appear to decrease the SIS_UTIL idleless level of the cpus
> enough to completely change the runqueue selection and migration behavior.
> 
> I fear that we hide a bad scheduler behavior under the rug by changing the
> idleless level of a specific workload pattern, while leaving the underlying
> root cause unfixed.
> 
> I'm currently working on a different approach: rate limit migrations.
> Basically, the idea is to detect when a task is migrated too often for its
> own good, and prevent the scheduler from migrating it for a short while. I
> get about 30% performance improvement with this approach as well (limit
> migration to 1 per 2ms window per task). I'll finish polishing my commit
> messages and send a series as RFC soon.
> 
> Thanks,
> 
> Mathieu
> 
> -- 
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com
>