| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-Id: <e873d01af6340976ba8d344caf824fd122d0a93d.1686554037.git.yu.c.chen@intel.com>
Date: Tue, 13 Jun 2023 00:19:13 +0800
From: Chen Yu <yu.c.chen@...el.com>
To: Peter Zijlstra <peterz@...radead.org>,
Vincent Guittot <vincent.guittot@...aro.org>,
Ingo Molnar <mingo@...hat.com>,
Juri Lelli <juri.lelli@...hat.com>
Cc: Tim Chen <tim.c.chen@...el.com>,
Mel Gorman <mgorman@...hsingularity.net>,
Dietmar Eggemann <dietmar.eggemann@....com>,
K Prateek Nayak <kprateek.nayak@....com>,
Abel Wu <wuyun.abel@...edance.com>,
"Gautham R . Shenoy" <gautham.shenoy@....com>,
Len Brown <len.brown@...el.com>,
Chen Yu <yu.chen.surf@...il.com>,
Yicong Yang <yangyicong@...ilicon.com>,
linux-kernel@...r.kernel.org, Chen Yu <yu.c.chen@...el.com>,
kernel test robot <lkp@...el.com>
Subject: [RFC PATCH 4/4] sched/fair: Throttle the busiest group scanning in idle load balance
Scanning the whole sched domain to find the busiest group is time costly
during newidle_balance(). Limit the scan depth of newidle_balance()
to only scan for a limited number of sched groups to find a relatively
busy group, and pull from it. The scanning depth is suggested by
the previous periodic load balance based on its overall utilization.
Tested on top of sched/core v6.4-rc1,
Sapphire Rapids with 2 x 56C/112T = 224 CPUs,
cpufreq governor is set to performance, and C6 is disabled.
Overall some improvements were noticed when the system is underloaded
from tbench/netperf, and no noticeable difference from hackbench/schbench.
And the percentage of newidle_balance() has dropped accordingly.
[netperf]
Launches $nr instances of:
netperf -4 -H 127.0.0.1 -t $work_mode -c -C -l 100 &
nr: 56, 112, 168, 224, 280, 336, 392, 448
work_mode: TCP_RR UDP_RR
throughput
=======
case load baseline(std%) compare%( std%)
TCP_RR 56-threads 1.00 ( 1.98) +18.45 ( 10.84)
TCP_RR 112-threads 1.00 ( 3.79) +0.92 ( 4.72)
TCP_RR 168-threads 1.00 ( 5.40) -0.09 ( 5.94)
TCP_RR 224-threads 1.00 ( 42.40) -1.37 ( 41.42)
TCP_RR 280-threads 1.00 ( 15.95) -0.30 ( 14.82)
TCP_RR 336-threads 1.00 ( 27.84) -0.08 ( 28.91)
TCP_RR 392-threads 1.00 ( 41.85) -0.56 ( 39.18)
TCP_RR 448-threads 1.00 ( 45.95) +1.62 ( 52.54)
UDP_RR 56-threads 1.00 ( 8.41) +4.86 ( 5.54)
UDP_RR 112-threads 1.00 ( 9.11) -0.68 ( 9.92)
UDP_RR 168-threads 1.00 ( 10.48) -0.15 ( 10.07)
UDP_RR 224-threads 1.00 ( 40.98) -3.80 ( 40.01)
UDP_RR 280-threads 1.00 ( 23.50) -0.53 ( 23.42)
UDP_RR 336-threads 1.00 ( 35.87) +0.38 ( 33.43)
UDP_RR 392-threads 1.00 ( 49.47) -0.27 ( 44.40)
UDP_RR 448-threads 1.00 ( 53.09) +1.81 ( 59.98)
[tbench]
tbench -t 100 $job 127.0.0.1
job: 56, 112, 168, 224, 280, 336, 392, 448
throughput
=========
loopback 56-threads 1.00 ( 1.12) +1.41 ( 0.43)
loopback 112-threads 1.00 ( 0.43) +0.30 ( 0.73)
loopback 168-threads 1.00 ( 6.88) -5.73 ( 7.74)
loopback 224-threads 1.00 ( 12.99) +31.32 ( 0.22)
loopback 280-threads 1.00 ( 0.38) -0.94 ( 0.99)
loopback 336-threads 1.00 ( 0.13) +0.06 ( 0.18)
loopback 392-threads 1.00 ( 0.06) -0.09 ( 0.16)
loopback 448-threads 1.00 ( 0.10) -0.13 ( 0.18)
[hackbench]
hackbench -g $job --$work_type --pipe -l 200000 -s 100 -f 28
and
hackbench -g $job --$work_type -l 200000 -s 100 -f 28
job: 1, 2, 4, 8
work_type: process threads
throughput
=========
case load baseline(std%) compare%( std%)
process-pipe 1-groups 1.00 ( 6.09) +2.61 ( 9.27)
process-pipe 2-groups 1.00 ( 7.15) +6.22 ( 5.59)
process-pipe 4-groups 1.00 ( 3.40) +2.01 ( 5.45)
process-pipe 8-groups 1.00 ( 0.44) -1.57 ( 0.70)
process-sockets 1-groups 1.00 ( 0.69) +0.86 ( 1.84)
process-sockets 2-groups 1.00 ( 5.04) -6.31 ( 0.60)
process-sockets 4-groups 1.00 ( 0.22) +0.01 ( 0.75)
process-sockets 8-groups 1.00 ( 0.49) +0.46 ( 0.49)
threads-pipe 1-groups 1.00 ( 1.96) -4.86 ( 6.90)
threads-pipe 2-groups 1.00 ( 3.02) +0.21 ( 2.72)
threads-pipe 4-groups 1.00 ( 4.83) -1.08 ( 6.41)
threads-pipe 8-groups 1.00 ( 3.86) +4.19 ( 3.82)
threads-sockets 1-groups 1.00 ( 2.20) +1.65 ( 1.85)
threads-sockets 2-groups 1.00 ( 3.09) -0.36 ( 2.14)
threads-sockets 4-groups 1.00 ( 0.99) -2.54 ( 1.86)
threads-sockets 8-groups 1.00 ( 0.27) -0.01 ( 0.79)
[schbench]
schbench -m $job -t 56 -r 30
job: 1, 2, 4, 8
3 iterations
99.0th latency
========
case load baseline(std%) compare%( std%)
normal 1-mthreads 1.00 ( 1.10) +0.88 ( 0.84)
normal 2-mthreads 1.00 ( 0.73) +0.00 ( 0.73)
normal 4-mthreads 1.00 ( 1.46) -0.60 ( 2.74)
normal 8-mthreads 1.00 ( 4.09) +1.08 ( 4.60)
Suggested-by: Tim Chen <tim.c.chen@...el.com>
Reported-by: kernel test robot <lkp@...el.com>
Signed-off-by: Chen Yu <yu.c.chen@...el.com>
---
kernel/sched/fair.c | 15 ++++++++++++++-
1 file changed, 14 insertions(+), 1 deletion(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index f999e838114e..272e6c224b96 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10168,7 +10168,12 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
struct sg_lb_stats *local = &sds->local_stat;
struct sg_lb_stats tmp_sgs;
unsigned long sum_util = 0;
- int sg_status = 0;
+ int sg_status = 0, nr_scan_ilb;
+ bool ilb_util_enabled = sched_feat(ILB_UTIL) && env->idle == CPU_NEWLY_IDLE &&
+ sd_share && READ_ONCE(sd_share->ilb_total_capacity);
+
+ if (ilb_util_enabled)
+ nr_scan_ilb = sd_share->ilb_nr_scan;
do {
struct sg_lb_stats *sgs = &tmp_sgs;
@@ -10186,6 +10191,14 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
+ if (ilb_util_enabled && --nr_scan_ilb <= 0) {
+ /* borrow the statistic of previous periodic load balance */
+ sds->total_load = READ_ONCE(sd_share->ilb_total_load);
+ sds->total_capacity = READ_ONCE(sd_share->ilb_total_capacity);
+
+ break;
+ }
+
if (local_group)
goto next_group;
--
2.25.1
Powered by blists - more mailing lists