Hi,
On 04/12/19 19:14, Srikar Dronamraju wrote:
> With commit 247f2f6f3c70 ("sched/core: Don't schedule threads on pre-empted
> vCPUs"), scheduler avoids preempted vCPUs to schedule tasks on wakeup.
> This leads to wrong choice of CPU, which in-turn leads to larger wakeup
> latencies. Eventually, it leads to performance regression in latency
> sensitive benchmarks like soltp, schbench etc.
>
> On Powerpc, vcpu_is_preempted only looks at yield_count. If the
> yield_count is odd, the vCPU is assumed to be preempted. However
> yield_count is increased whenever LPAR enters CEDE state. So any CPU
> that has entered CEDE state is assumed to be preempted.
>
> Even if vCPU of dedicated LPAR is preempted/donated, it should have
> right of first-use since they are suppose to own the vCPU.
>
> On a Power9 System with 32 cores
> # lscpu
> Architecture: ppc64le
> Byte Order: Little Endian
> CPU(s): 128
> On-line CPU(s) list: 0-127
> Thread(s) per core: 8
> Core(s) per socket: 1
> Socket(s): 16
> NUMA node(s): 2
> Model: 2.2 (pvr 004e 0202)
> Model name: POWER9 (architected), altivec supported
> Hypervisor vendor: pHyp
> Virtualization type: para
> L1d cache: 32K
> L1i cache: 32K
> L2 cache: 512K
> L3 cache: 10240K
> NUMA node0 CPU(s): 0-63
> NUMA node1 CPU(s): 64-127
>
>
> # perf stat -a -r 5 ./schbench
> v5.4 v5.4 + patch
> Latency percentiles (usec) Latency percentiles (usec)
> 49.0000th: 47 50.0000th: 33
> 74.0000th: 64 75.0000th: 44
> 89.0000th: 76 90.0000th: 50
> 94.0000th: 83 95.0000th: 53
> *98.0000th: 103 *99.0000th: 57
> 98.5000th: 2124 99.5000th: 59
> 98.9000th: 7976 99.9000th: 83
> min=-1, max=10519 min=0, max=117
> Latency percentiles (usec) Latency percentiles (usec)
> 49.0000th: 45 50.0000th: 34
> 74.0000th: 61 75.0000th: 45
> 89.0000th: 70 90.0000th: 52
> 94.0000th: 77 95.0000th: 56
> *98.0000th: 504 *99.0000th: 62
> 98.5000th: 4012 99.5000th: 64
> 98.9000th: 8168 99.9000th: 79
> min=-1, max=14500 min=0, max=123
> Latency percentiles (usec) Latency percentiles (usec)
> 49.0000th: 48 50.0000th: 35
> 74.0000th: 65 75.0000th: 47
> 89.0000th: 76 90.0000th: 55
> 94.0000th: 82 95.0000th: 59
> *98.0000th: 1098 *99.0000th: 67
> 98.5000th: 3988 99.5000th: 71
> 98.9000th: 9360 99.9000th: 98
> min=-1, max=19283 min=0, max=137
> Latency percentiles (usec) Latency percentiles (usec)
> 49.0000th: 46 50.0000th: 35
> 74.0000th: 63 75.0000th: 46
> 89.0000th: 73 90.0000th: 53
> 94.0000th: 78 95.0000th: 57
> *98.0000th: 113 *99.0000th: 63
> 98.5000th: 2316 99.5000th: 65
> 98.9000th: 7704 99.9000th: 83
> min=-1, max=17976 min=0, max=139
> Latency percentiles (usec) Latency percentiles (usec)
> 49.0000th: 46 50.0000th: 34
> 74.0000th: 62 75.0000th: 46
> 89.0000th: 73 90.0000th: 53
> 94.0000th: 79 95.0000th: 57
> *98.0000th: 97 *99.0000th: 64
> 98.5000th: 1398 99.5000th: 70
> 98.9000th: 8136 99.9000th: 100
> min=-1, max=10008 min=0, max=142
>
> Performance counter stats for 'system wide' (4 runs):
>
> context-switches 42,604 ( +- 0.87% ) 45,397 ( +- 0.25% )
> cpu-migrations 0,195 ( +- 2.70% ) 230 ( +- 7.23% )
> page-faults 16,783 ( +- 14.87% ) 16,781 ( +- 9.77% )
>
> Waiman Long suggested using static_keys.
>
> Reported-by: Parth Shah <pa...@linux.ibm.com>
> Reported-by: Ihor Pasichnyk <ihor.pasich...@ibm.com>
> Cc: Parth Shah <pa...@linux.ibm.com>
> Cc: Ihor Pasichnyk <ihor.pasich...@ibm.com>
> Cc: Juri Lelli <juri.le...@redhat.com>
> Cc: Waiman Long <long...@redhat.com>
> Signed-off-by: Srikar Dronamraju <sri...@linux.vnet.ibm.com>
> ---
I could reproduce the performace regression and can confirm that this
patch fixes it.
Tested-by: Juri Lelli <juri.le...@redhat.com>
Thanks,
Juri
