Re: [rcu] 5057f55e543: -23.5% qperf.udp.recv_bw
From: Paul E. McKenney
Date: Tue Jun 03 2014 - 11:44:59 EST
On Tue, Jun 03, 2014 at 06:08:41PM +0800, Jet Chen wrote:
> Hi Paul,
>
> FYI, we noticed the below changes on
>
> git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git rcu/fixes
> commit 5057f55e543b7859cfd26bc281291795eac93f8a ("rcu: Bind RCU grace-period kthreads if NO_HZ_FULL")
My guess would be that some of these workloads generated enough callbacks
that binding all the rcuo callback-offloading kthreads to CPU 0 resulted
in a bottleneck. If that was the case, CPU 0 would often hit 100%
CPU utilization, and there would be more wait time on other CPUs because
callback execution was delayed.
Does that match what you are seeing?
Thanx, Paul
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.127e+09 ~ 0% -23.5% 1.628e+09 ~ 4% bens/qperf/600s
> 2.127e+09 ~ 0% -23.5% 1.628e+09 ~ 4% TOTAL qperf.udp.recv_bw
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.128e+09 ~ 0% -23.3% 1.633e+09 ~ 4% bens/qperf/600s
> 2.128e+09 ~ 0% -23.3% 1.633e+09 ~ 4% TOTAL qperf.udp.send_bw
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.101e+10 ~ 2% -18.7% 1.707e+10 ~ 2% bens/iperf/300s-tcp
> 2.101e+10 ~ 2% -18.7% 1.707e+10 ~ 2% TOTAL iperf.tcp.sender.bps
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.101e+10 ~ 2% -18.7% 1.707e+10 ~ 2% bens/iperf/300s-tcp
> 2.101e+10 ~ 2% -18.7% 1.707e+10 ~ 2% TOTAL iperf.tcp.receiver.bps
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 1.331e+09 ~ 2% -5.8% 1.255e+09 ~ 2% bens/qperf/600s
> 2.4e+09 ~ 6% -30.4% 1.671e+09 ~12% brickland3/qperf/600s
> 2.384e+09 ~ 7% -12.1% 2.096e+09 ~ 3% lkp-sb03/qperf/600s
> 6.115e+09 ~ 5% -17.9% 5.022e+09 ~ 6% TOTAL qperf.sctp.bw
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.83e+09 ~ 1% -12.5% 2.476e+09 ~ 3% bens/qperf/600s
> 2.83e+09 ~ 1% -12.5% 2.476e+09 ~ 3% TOTAL qperf.tcp.bw
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 2.272e+08 ~ 1% -13.3% 1.97e+08 ~ 2% bens/qperf/600s
> 2.272e+08 ~ 1% -13.3% 1.97e+08 ~ 2% TOTAL proc-vmstat.pgalloc_dma32
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 53062 ~ 2% -35.1% 34464 ~ 3% bens/qperf/600s
> 109531 ~13% +46.9% 160928 ~ 5% brickland3/qperf/600s
> 67902 ~ 1% +13.8% 77302 ~ 3% lkp-sb03/qperf/600s
> 230496 ~ 7% +18.3% 272694 ~ 4% TOTAL softirqs.RCU
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 80344 ~ 1% -26.2% 59325 ~ 2% bens/qperf/600s
> 80344 ~ 1% -26.2% 59325 ~ 2% TOTAL softirqs.SCHED
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 1036 ~ 4% -17.6% 853 ~ 4% brickland3/qperf/600s
> 1036 ~ 4% -17.6% 853 ~ 4% TOTAL proc-vmstat.nr_page_table_pages
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 48.12 ~ 0% -11.7% 42.46 ~ 6% brickland3/qperf/600s
> 48.12 ~ 0% -11.7% 42.46 ~ 6% TOTAL turbostat.%pc2
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 74689352 ~ 1% -13.3% 64771743 ~ 2% bens/qperf/600s
> 74689352 ~ 1% -13.3% 64771743 ~ 2% TOTAL proc-vmstat.pgalloc_normal
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 3.019e+08 ~ 1% -13.3% 2.618e+08 ~ 2% bens/qperf/600s
> 3.019e+08 ~ 1% -13.3% 2.618e+08 ~ 2% TOTAL proc-vmstat.pgfree
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 23538414 ~ 0% -12.9% 20506157 ~ 2% bens/qperf/600s
> 23538414 ~ 0% -12.9% 20506157 ~ 2% TOTAL proc-vmstat.numa_local
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 23538414 ~ 0% -12.9% 20506157 ~ 2% bens/qperf/600s
> 23538414 ~ 0% -12.9% 20506157 ~ 2% TOTAL proc-vmstat.numa_hit
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 12789 ~ 1% -10.9% 11391 ~ 2% bens/qperf/600s
> 12789 ~ 1% -10.9% 11391 ~ 2% TOTAL softirqs.HRTIMER
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 481253 ~ 0% -8.9% 438624 ~ 0% bens/qperf/600s
> 481253 ~ 0% -8.9% 438624 ~ 0% TOTAL softirqs.TIMER
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 1297 ~33% +565.9% 8640 ~ 7% bens/iperf/300s-tcp
> 2788 ~ 3% +588.8% 19204 ~ 4% bens/qperf/600s
> 1191 ~ 5% +1200.9% 15493 ~ 4% brickland3/qperf/600s
> 1135 ~26% +1195.9% 14709 ~ 4% lkp-sb03/qperf/600s
> 6411 ~13% +805.3% 58047 ~ 4% TOTAL time.involuntary_context_switches
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 72398 ~ 1% -5.4% 68503 ~ 0% bens/qperf/600s
> 8789 ~ 4% +22.3% 10749 ~15% lkp-sb03/qperf/600s
> 81187 ~ 1% -2.4% 79253 ~ 2% TOTAL vmstat.system.in
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 141174 ~ 1% -5.4% 133551 ~ 0% bens/qperf/600s
> 143982 ~ 1% -4.4% 137600 ~ 0% brickland3/qperf/600s
> 285156 ~ 1% -4.9% 271152 ~ 0% TOTAL vmstat.system.cs
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 42351859 ~ 0% -5.3% 40114932 ~ 0% bens/qperf/600s
> 43015383 ~ 1% -4.4% 41143092 ~ 0% brickland3/qperf/600s
> 85367242 ~ 1% -4.8% 81258025 ~ 0% TOTAL time.voluntary_context_switches
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 146 ~ 0% -2.2% 143 ~ 0% bens/qperf/600s
> 147 ~ 1% -4.8% 140 ~ 1% brickland3/qperf/600s
> 293 ~ 0% -3.5% 283 ~ 0% TOTAL time.percent_of_cpu_this_job_got
>
> 71a9b26963f8c2d 5057f55e543b7859cfd26bc28
> --------------- -------------------------
> 872 ~ 0% -2.3% 853 ~ 0% bens/qperf/600s
> 874 ~ 1% -4.6% 834 ~ 1% brickland3/qperf/600s
> 1747 ~ 0% -3.4% 1687 ~ 0% TOTAL time.system_time
>
>
> Legend:
> ~XX% - stddev percent
> [+-]XX% - change percent
>
>
>
>
> Disclaimer:
> Results have been estimated based on internal Intel analysis and are provided
> for informational purposes only. Any difference in system hardware or software
> design or configuration may affect actual performance.
>
> Thanks,
> Jet
>
>
> echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
> echo performance > /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
> echo performance > /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor
> echo performance > /sys/devices/system/cpu/cpu3/cpufreq/scaling_governor
> ./runtest.py unlink1 32 1 2 3 4
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/