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

From: Mathieu Desnoyers
Date: Wed Aug 23 2023 - 10:17:03 EST


On 8/23/23 10:05, Mathieu Desnoyers wrote:
On 8/23/23 02:08, 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. After this change, the cost of accessing tg->load_avg is greatly
reduced and performance improved. Detailed test results below.

By applying your patch on top of my patchset at:

https://lore.kernel.org/lkml/20230822113133.643238-1-mathieu.desnoyers@xxxxxxxxxxxx/

The combined hackbench results look very promising:

(hackbench -g 32 -f 20 --threads --pipe -l 480000 -s 100)
(192 cores AMD EPYC 9654 96-Core Processor (over 2 sockets), with hyperthreading)

Baseline:                                       49s
With L2-ttwu-queue-skip:                        34s (30% speedup)
With L2-ttwu-queue-skip + ratelimit-load-avg:   26s (46% speedup)

Here is an additional interesting data point:

With only ratelimit-load-avg patch: 32s (35% speedup)

So each series appear to address a different scalability issue, and
combining both seems worthwhile, at least from the point of view of
this specific benchmark on this hardware.

I'm looking forward to see numbers for other benchmarks and hardware.

Thanks,

Mathieu


Feel free to apply my:

Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
Tested-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>

Thanks Aaron!

Mathieu


==============================
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@xxxxxxxxx>
Suggested-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
Signed-off-by: Aaron Lu <aaron.lu@xxxxxxxxx>
Reviewed-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
---
  kernel/sched/fair.c  | 13 ++++++++++++-
  kernel/sched/sched.h |  1 +
  2 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index c28206499a3d..a5462d1fcc48 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3664,7 +3664,8 @@ static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
   */
  static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
  {
-    long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
+    long delta;
+    u64 now;
      /*
       * No need to update load_avg for root_task_group as it is not used.
@@ -3672,9 +3673,19 @@ static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
      if (cfs_rq->tg == &root_task_group)
          return;
+    /*
+     * For migration heavy workload, access to tg->load_avg can be
+     * unbound. Limit the update rate to at most once per ms.
+     */
+    now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
+    if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
+        return;
+
+    delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
      if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
          atomic_long_add(delta, &cfs_rq->tg->load_avg);
          cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
+        cfs_rq->last_update_tg_load_avg = now;
      }
  }
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 6a8b7b9ed089..52ee7027def9 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -593,6 +593,7 @@ struct cfs_rq {
      } removed;
  #ifdef CONFIG_FAIR_GROUP_SCHED
+    u64            last_update_tg_load_avg;
      unsigned long        tg_load_avg_contrib;
      long            propagate;
      long            prop_runnable_sum;


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