Re: [PATCH v4 00/10] steal tasks to improve CPU utilization

[Steven Sistare <steven.sistare@oracle.com>]

On 12/10/2018 11:10 AM, Vincent Guittot wrote:
> Hi Steven,
> 
> On Thu, 6 Dec 2018 at 22:38, Steve Sistare <steven.sistare@oracle.com> wrote:
>>
>> When a CPU has no more CFS tasks to run, and idle_balance() fails to
>> find a task, then attempt to steal a task from an overloaded CPU in the
>> same LLC. Maintain and use a bitmap of overloaded CPUs to efficiently
>> identify candidates.  To minimize search time, steal the first migratable
>> task that is found when the bitmap is traversed.  For fairness, search
>> for migratable tasks on an overloaded CPU in order of next to run.
>>
>> This simple stealing yields a higher CPU utilization than idle_balance()
>> alone, because the search is cheap, so it may be called every time the CPU
>> is about to go idle.  idle_balance() does more work because it searches
>> widely for the busiest queue, so to limit its CPU consumption, it declines
>> to search if the system is too busy.  Simple stealing does not offload the
>> globally busiest queue, but it is much better than running nothing at all.
>>
>> The bitmap of overloaded CPUs is a new type of sparse bitmap, designed to
>> reduce cache contention vs the usual bitmap when many threads concurrently
>> set, clear, and visit elements.
>>
>> Patch 1 defines the sparsemask type and its operations.
>>
>> Patches 2, 3, and 4 implement the bitmap of overloaded CPUs.
>>
>> Patches 5 and 6 refactor existing code for a cleaner merge of later
>>   patches.
>>
>> Patches 7 and 8 implement task stealing using the overloaded CPUs bitmap.
>>
>> Patch 9 disables stealing on systems with more than 2 NUMA nodes for the
>> time being because of performance regressions that are not due to stealing
>> per-se.  See the patch description for details.
>>
>> Patch 10 adds schedstats for comparing the new behavior to the old, and
>>   provided as a convenience for developers only, not for integration.
>>
>> The patch series is based on kernel 4.20.0-rc1.  It compiles, boots, and
>> runs with/without each of CONFIG_SCHED_SMT, CONFIG_SMP, CONFIG_SCHED_DEBUG,
>> and CONFIG_PREEMPT.  It runs without error with CONFIG_DEBUG_PREEMPT +
>> CONFIG_SLUB_DEBUG + CONFIG_DEBUG_PAGEALLOC + CONFIG_DEBUG_MUTEXES +
>> CONFIG_DEBUG_SPINLOCK + CONFIG_DEBUG_ATOMIC_SLEEP.  CPU hot plug and CPU
>> bandwidth control were tested.
>>
>> Stealing improves utilization with only a modest CPU overhead in scheduler
>> code.  In the following experiment, hackbench is run with varying numbers
>> of groups (40 tasks per group), and the delta in /proc/schedstat is shown
>> for each run, averaged per CPU, augmented with these non-standard stats:
>>
>>   %find - percent of time spent in old and new functions that search for
>>     idle CPUs and tasks to steal and set the overloaded CPUs bitmap.
>>
>>   steal - number of times a task is stolen from another CPU.
>>
>> X6-2: 1 socket * 10 cores * 2 hyperthreads = 20 CPUs
>> Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz
>> hackbench <grps> process 100000
>> sched_wakeup_granularity_ns=15000000
>>
>>   baseline
>>   grps  time  %busy  slice   sched   idle     wake %find  steal
>>   1    8.084  75.02   0.10  105476  46291    59183  0.31      0
>>   2   13.892  85.33   0.10  190225  70958   119264  0.45      0
>>   3   19.668  89.04   0.10  263896  87047   176850  0.49      0
>>   4   25.279  91.28   0.10  322171  94691   227474  0.51      0
>>   8   47.832  94.86   0.09  630636 144141   486322  0.56      0
>>
>>   new
>>   grps  time  %busy  slice   sched   idle     wake %find  steal  %speedup
>>   1    5.938  96.80   0.24   31255   7190    24061  0.63   7433  36.1
>>   2   11.491  99.23   0.16   74097   4578    69512  0.84  19463  20.9
>>   3   16.987  99.66   0.15  115824   1985   113826  0.77  24707  15.8
>>   4   22.504  99.80   0.14  167188   2385   164786  0.75  29353  12.3
>>   8   44.441  99.86   0.11  389153   1616   387401  0.67  38190   7.6
>>
>> Elapsed time improves by 8 to 36%, and CPU busy utilization is up
>> by 5 to 22% hitting 99% for 2 or more groups (80 or more tasks).
>> The cost is at most 0.4% more find time.
> 
> I have run some hackbench tests on my hikey arm64 octo cores with your
> patchset. My original intent was to send a tested-by but I have some
> performances regressions.
> This hikey is the smp one and not the asymetric hikey960 that Valentin
> used for his tests
> The sched domain topology is
> domain-0: span=0-3 level=MC  and domain-0: span=4-7 level=MC
> domain-1: span=0-7 level=DIE
> 
> I have run 12 times hackbench -g $j -P -l 2000 with j equals to 1 2 3 4 8
> 
> grps  time
> 1      1.396
> 2      2.699
> 3      3.617
> 4      4.498
> 8      7.721
> 
> Then after disabling STEAL in sched_feature with echo NO_STEAL >
> /sys/kernel/debug/sched_features , the results become:
> grps  time
> 1      1.217
> 2      1.973
> 3      2.855
> 4      3.932
> 8      7.674
> 
> I haven't looked in details about some possible reasons of such
> difference yet and haven't collected the stats that you added with
> patch 10.
> Have you got a script to collect and post process them ?
> 
> Regards,
> Vincent

Thanks Vincent.  What is the value of /proc/sys/kernel/sched_wakeup_granularity_ns?
Try 15000000.  Your 8-core system is heavily overloaded with 40 * groups tasks,
and I suspect preemptions are killing performance.

I have a python script to post-process schedstat files, but it does many things
and is large and I am not ready to share it.  I can write a short bash script if
that would help.

- Steve