Re: [PATCH -next] sched/fair: inline cpu_util_without and cpu_util to improve performance

[Li Zetao <lizetao1@huawei.com>]

Hi,

在 2024/7/24 18:53, Peter Zijlstra 写道:
> On Tue, Jul 23, 2024 at 03:36:07PM +0800, Li Zetao wrote:
>> The commit 3eb6d6ececca ("sched/fair: Refactor CPU utilization functions")
>> refactored cpu_util_without and cpu_util functions. Since the size of
>> cpu_util function has increased, the inline cpu_util is dropped. This had
>> a negative impact on performance, in the scenario of updating
>> sched_group's statistics, cpu_util_without and cpu_util functions are on
>> the hotspot path.
>>
>> Inlining cpu_util_without and cpu_util functions have been shown to
>> significantly improve performance in lmbench as follow:
>>
>>    Machine: Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz
>>                       before          after          diff
>>    fork+exit          317.0625        303.6667       -4.22%
>>    fork+execve        1482.5000       1407.0000      -5.09%
>>    fork+/bin/sh       2096.0000       2020.3333      -3.61%
> 
> That's quite significant. Did you look at what exactly was causing this?

In the fork scene, in order to select a CPU for new task, there is the 
following hotspot path:

   sched_balance_find_dst_cpu() ->
     sched_balance_find_dst_group() ->
       update_sg_wakeup_stats() -> cpu_util_without() -> cpu_util()

When the system has a large number of CPUs, in order to calculate the 
util of sched_group, the cpu_util() will be called many times.
> 
>> This patch introduces inlining to cpu_util_without and cpu_util functions.
>> While this increases the size of kernel/sched/fair.o, the performance
>> gains in critical workloads make this an acceptable trade-off.
>>
>> Size comparison before and after patch:
>>       text	   data	    bss	    dec	    hex	filename
>>     0x1264a	 0x1506	   0xb0	  80896	  13c00	kernel/sched/fair.o.before
>>     0x12672	 0x14fe	   0xb0	  80928	  13c20	kernel/sched/fair.o.after
>>
>> Signed-off-by: Zhang Qiao <zhangqiao22@huawei.com>
>> Signed-off-by: Li Zetao <lizetao1@huawei.com>
> 
> This SoB chain is not valid. Please review the documentation we have
> on this.

I will fix this in v2 version.

> 
>> ---
>>   kernel/sched/fair.c | 4 ++--
>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 5904405ffc59..677b78fa65b6 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -7706,7 +7706,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
>>    *
>>    * Return: (Boosted) (estimated) utilization for the specified CPU.
>>    */
>> -static unsigned long
>> +static __always_inline unsigned long
>>   cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
>>   {
>>   	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
>> @@ -7794,7 +7794,7 @@ unsigned long cpu_util_cfs_boost(int cpu)
>>    * utilization of the specified task, whenever the task is currently
>>    * contributing to the CPU utilization.
>>    */
>> -static unsigned long cpu_util_without(int cpu, struct task_struct *p)
>> +static __always_inline unsigned long cpu_util_without(int cpu, struct task_struct *p)
>>   {
>>   	/* Task has no contribution or is new */
>>   	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
>> -- 
>> 2.34.1
>>
> 

Li Zetao
Best regard