Re: [PATCH v5 08/24] sched: Introduce per memory space current virtual cpu id

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 2022-11-08 08:04, Peter Zijlstra wrote:
> On Thu, Nov 03, 2022 at 04:03:43PM -0400, Mathieu Desnoyers wrote:
> 
>> The credit goes to Paul Turner (Google) for the vcpu_id idea. This
>> feature is implemented based on the discussions with Paul Turner and
>> Peter Oskolkov (Google), but I took the liberty to implement scheduler
>> fast-path optimizations and my own NUMA-awareness scheme. The rumor has
>> it that Google have been running a rseq vcpu_id extension internally at
>> Google in production for a year. The tcmalloc source code indeed has
>> comments hinting at a vcpu_id prototype extension to the rseq system
>> call [1].
> 
> Re NUMA thing -- that means that on a 512 node system a single threaded
> task can still observe 512 separate vcpu-ids, right?

Yes, that's correct.

> 
> Also, said space won't be dense.

Indeed, this can be inefficient if the data structure within the 
single-threaded task is not NUMA-aware *and* that task is free to bounce 
all over the 512 numa nodes.

> 
> The main selling point of the whole vcpu-id scheme was that the id space
> is dense and not larger than min(nr_cpus, nr_threads), which then gives
> useful properties.
> 
> But I'm not at all seeing how the NUMA thing preserves that.

If a userspace per-vcpu data structure is implemented with NUMA-local 
allocations, then it becomes really interesting to guarantee that the 
per-vcpu-id accesses are always numa-local for performance reasons.

If a userspace per-vcpu data structure is not numa-aware, then we have 
two scenarios:

A) The cpuset/sched affinity under which it runs pins it to a set of 
cores belonging to a specific NUMA node. In this case, even with 
numa-aware vcpu id allocation, the ids will stay as close to 0 as if not 
numa-aware.

B) No specific cpuset/sched affinity set, which means the task is free 
to bounce all over. In this case I agree that having the indexing 
numa-aware, but the per-vcpu data structure not numa-aware, is inefficient.

I wonder whether scenarios with 512 nodes systems, with containers using 
few cores, but without using cpusets/sched affinity to pin the workload 
to specific numa nodes is a workload we should optimize for ? It looks 
like the lack of numa locality due to lack of allowed cores restriction 
is a userspace configuration issue.

We also must keep in mind that we can expect a single task to load a mix 
of executable/shared libraries where some pieces may be numa-aware, and 
others may not. This means we should ideally support a numa-aware 
vcpu-id allocation scheme and non-numa-aware vcpu-id allocation scheme 
within the same task.

This could be achieved by exposing two struct rseq fields rather than 
one, e.g.:

vm_vcpu_id -> flat indexing, not numa-aware.
vm_numa_vcpu_id -> numa-aware vcpu id indexing.

This would allow data structures that are inherently numa-aware to 
benefit from numa-locality, without hurting non-numa-aware data structures.

> 
> Also; given the utter mind-bendiness of the NUMA thing; should it go
> into it's own patch; introduce the regular plain old vcpu first, and
> then add things to it -- that also allows pushing those weird cpumask
> ops you've created later into the series.

Good idea. I can do that once we agree on the way forward for flat vs 
numa-aware vcpu-id rseq fields.

Thanks,

Mathieu


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