Re: [tip:sched/core] sched/x86: Optimize switch_mm() for multi-threaded workloads

[Joe Mario <jmario@redhat.com>]

On 08/02/2013 05:12 AM, Ingo Molnar wrote:
> * tip-bot for Rik van Riel <tipbot@zytor.com> wrote:
>
>> Commit-ID:  8f898fbbe5ee5e20a77c4074472a1fd088dc47d1
>> Gitweb:     http://git.kernel.org/tip/8f898fbbe5ee5e20a77c4074472a1fd088dc47d1
>> Author:     Rik van Riel <riel@redhat.com>
>> AuthorDate: Wed, 31 Jul 2013 22:14:21 -0400
>> Committer:  Ingo Molnar <mingo@kernel.org>
>> CommitDate: Thu, 1 Aug 2013 09:10:26 +0200
>>
>> sched/x86: Optimize switch_mm() for multi-threaded workloads
>>
>> Dick Fowles, Don Zickus and Joe Mario have been working on
>> improvements to perf, and noticed heavy cache line contention
>> on the mm_cpumask, running linpack on a 60 core / 120 thread
>> system.
>>
>> The cause turned out to be unnecessary atomic accesses to the
>> mm_cpumask. When in lazy TLB mode, the CPU is only removed from
>> the mm_cpumask if there is a TLB flush event.
>>
>> Most of the time, no such TLB flush happens, and the kernel
>> skips the TLB reload. It can also skip the atomic memory
>> set & test.
>>
>> Here is a summary of Joe's test results:
>>
>>   * The __schedule function dropped from 24% of all program cycles down
>>     to 5.5%.
>>
>>   * The cacheline contention/hotness for accesses to that bitmask went
>>     from being the 1st/2nd hottest - down to the 84th hottest (0.3% of
>>     all shared misses which is now quite cold)
>>
>>   * The average load latency for the bit-test-n-set instruction in
>>     __schedule dropped from 10k-15k cycles down to an average of 600 cycles.
>>
>>   * The linpack program results improved from 133 GFlops to 144 GFlops.
>>     Peak GFlops rose from 133 to 153.
>>
>> Reported-by: Don Zickus <dzickus@redhat.com>
>> Reported-by: Joe Mario <jmario@redhat.com>
>> Tested-by: Joe Mario <jmario@redhat.com>
>> Signed-off-by: Rik van Riel <riel@redhat.com>
>> Reviewed-by: Paul Turner <pjt@google.com>
>> Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
>> Link: http://lkml.kernel.org/r/20130731221421.616d3d20@annuminas.surriel.com
>> [ Made the comments consistent around the modified code. ]
>> Signed-off-by: Ingo Molnar <mingo@kernel.org>
>> +	  else {
>>   		this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
>>   		BUG_ON(this_cpu_read(cpu_tlbstate.active_mm) != next);
>>   
>> -		if (!cpumask_test_and_set_cpu(cpu, mm_cpumask(next))) {
>> +		if (!cpumask_test_cpu(cpu, mm_cpumask(next))) {
>> +			/*
>> +			 * On established mms, the mm_cpumask is only changed
>> +			 * from irq context, from ptep_clear_flush() while in
>> +			 * lazy tlb mode, and here. Irqs are blocked during
>> +			 * schedule, protecting us from simultaneous changes.
>> +			 */
>> +			cpumask_set_cpu(cpu, mm_cpumask(next));
> Note, I marked this for v3.12 with no -stable backport tag as it's not a
> regression fix.
>
> Nevertheless if it's a real issue in production (and +20% of linpack
> performance is certainly significant)

The cacheline contention on the mm->cpu_vm_mask_va cpulist bitmask 
skyrockets on numa systems as soon as the number of threads exceeds the 
available cpus - and it's all from the locked bts instruction in 
cpumask_test_and_set_cpu().
The OMP version of Linpack uses clone() with the CLONE_VM flag set, so 
all the clones are tugging at the same mm->cpu_vm_mask_va.
With Rik's patch, the number of accesses tomm->cpu_vm_mask_va plummets.

> feel free to forward it to -stable
> once this hits Linus's tree in the v3.12 merge window - by that time the
> patch will be reasonably well tested and it's a relatively simple change.
>
> Thanks,
>
> 	Ingo