Re: [PATCH v2 1/1] sched/fair: ratelimit update to tg->load_avg

[Chen Yu <yu.c.chen@intel.com>]

On 2023-09-12 at 14:58:08 +0800, Aaron Lu wrote:
> When using sysbench to benchmark Postgres in a single docker instance
> with sysbench's nr_threads set to nr_cpu, it is observed there are times
> update_cfs_group() and update_load_avg() shows noticeable overhead on
> a 2sockets/112core/224cpu Intel Sapphire Rapids(SPR):
> 
>     13.75%    13.74%  [kernel.vmlinux]           [k] update_cfs_group
>     10.63%    10.04%  [kernel.vmlinux]           [k] update_load_avg
> 
> Annotate shows the cycles are mostly spent on accessing tg->load_avg
> with update_load_avg() being the write side and update_cfs_group() being
> the read side. tg->load_avg is per task group and when different tasks
> of the same taskgroup running on different CPUs frequently access
> tg->load_avg, it can be heavily contended.
> 
> E.g. when running postgres_sysbench on a 2sockets/112cores/224cpus Intel
> Sappire Rapids, during a 5s window, the wakeup number is 14millions and
> migration number is 11millions and with each migration, the task's load
> will transfer from src cfs_rq to target cfs_rq and each change involves
> an update to tg->load_avg. Since the workload can trigger as many wakeups
> and migrations, the access(both read and write) to tg->load_avg can be
> unbound. As a result, the two mentioned functions showed noticeable
> overhead. With netperf/nr_client=nr_cpu/UDP_RR, the problem is worse:
> during a 5s window, wakeup number is 21millions and migration number is
> 14millions; update_cfs_group() costs ~25% and update_load_avg() costs ~16%.
> 
> Reduce the overhead by limiting updates to tg->load_avg to at most once
> per ms. The update frequency is a tradeoff between tracking accuracy and
> overhead. 1ms is chosen because PELT window is roughly 1ms and it
> delivered good results for the tests that I've done. After this change,
> the cost of accessing tg->load_avg is greatly reduced and performance
> improved. Detailed test results below.
> 
> ==============================
> postgres_sysbench on SPR:
> 25%
> base:   42382±19.8%
> patch:  50174±9.5%  (noise)
> 
> 50%
> base:   67626±1.3%
> patch:  67365±3.1%  (noise)
> 
> 75%
> base:   100216±1.2%
> patch:  112470±0.1% +12.2%
> 
> 100%
> base:    93671±0.4%
> patch:  113563±0.2% +21.2%
> 
> ==============================
> hackbench on ICL:
> group=1
> base:    114912±5.2%
> patch:   117857±2.5%  (noise)
> 
> group=4
> base:    359902±1.6%
> patch:   361685±2.7%  (noise)
> 
> group=8
> base:    461070±0.8%
> patch:   491713±0.3% +6.6%
> 
> group=16
> base:    309032±5.0%
> patch:   378337±1.3% +22.4%
> 
> =============================
> hackbench on SPR:
> group=1
> base:    100768±2.9%
> patch:   103134±2.9%  (noise)
> 
> group=4
> base:    413830±12.5%
> patch:   378660±16.6% (noise)
> 
> group=8
> base:    436124±0.6%
> patch:   490787±3.2% +12.5%
> 
> group=16
> base:    457730±3.2%
> patch:   680452±1.3% +48.8%
> 
> ============================
> netperf/udp_rr on ICL
> 25%
> base:    114413±0.1%
> patch:   115111±0.0% +0.6%
> 
> 50%
> base:    86803±0.5%
> patch:   86611±0.0%  (noise)
> 
> 75%
> base:    35959±5.3%
> patch:   49801±0.6% +38.5%
> 
> 100%
> base:    61951±6.4%
> patch:   70224±0.8% +13.4%
> 
> ===========================
> netperf/udp_rr on SPR
> 25%
> base:   104954±1.3%
> patch:  107312±2.8%  (noise)
> 
> 50%
> base:    55394±4.6%
> patch:   54940±7.4%  (noise)
> 
> 75%
> base:    13779±3.1%
> patch:   36105±1.1% +162%
> 
> 100%
> base:     9703±3.7%
> patch:   28011±0.2% +189%
> 
> ==============================================
> netperf/tcp_stream on ICL (all in noise range)
> 25%
> base:    43092±0.1%
> patch:   42891±0.5%
> 
> 50%
> base:    19278±14.9%
> patch:   22369±7.2%
> 
> 75%
> base:    16822±3.0%
> patch:   17086±2.3%
> 
> 100%
> base:    18216±0.6%
> patch:   18078±2.9%
> 
> ===============================================
> netperf/tcp_stream on SPR (all in noise range)
> 25%
> base:    34491±0.3%
> patch:   34886±0.5%
> 
> 50%
> base:    19278±14.9%
> patch:   22369±7.2%
> 
> 75%
> base:    16822±3.0%
> patch:   17086±2.3%
> 
> 100%
> base:    18216±0.6%
> patch:   18078±2.9%
> 
> Reported-by: Nitin Tekchandani <nitin.tekchandani@intel.com>
> Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
> Signed-off-by: Aaron Lu <aaron.lu@intel.com>
> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
> Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Tested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Reviewed-by: David Vernet <void@manifault.com>
> Tested-by: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
> ---
>

Since we know that this patch brings good improvement for netperf,
hackbench, I did some further verification on tbench/schbench on Ice Lake
Xeon Platinum 8360Y, and it reports good result too:

schbench
========
case            	load    	baseline(std%)	compare%( std%)
normal          	1-mthreads	 1.00 (  1.70)	 +0.00 (  0.00)
normal          	2-mthreads	 1.00 (  2.32)	 -0.62 (  5.24)
normal          	4-mthreads	 1.00 (  3.17)	 -1.86 (  3.11)

tbench
======
case            	load    	baseline(std%)	compare%( std%)
loopback        	36-threads	 1.00 (  2.80)	 +1.85 (  0.45)
loopback        	72-threads	 1.00 (  0.27)	 -0.20 (  0.51)
loopback        	108-threads	 1.00 (  0.06)	+21.92 (  0.10)
loopback        	144-threads	 1.00 (  1.47)	+28.42 (  0.11)

Tested-by: Chen Yu <yu.c.chen@intel.com>

thanks,
Chenyu