lkml.org 
[lkml]   [2022]   [Feb]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] irqchip/gic-v3: use dsb(ishst) to synchronize data to smp before issuing ipi
Date
>> +	dsb(ishst);
>>
>> for_each_cpu(cpu, mask) {
>> u64 cluster_id = MPIDR_TO_SGI_CLUSTER_ID(cpu_logical_map(cpu));
>
> I'm not opposed to that change, but I'm pretty curious whether this
> makes
> any visible difference in practice. Could you measure the effect of this
> change
> for any sort of IPI heavy workload?
>
> Thanks,
>
> M.

In practice, at least I don't see much difference on the hardware I am
using. So the result probably depends on the implementaion of the real
hardwares.

I wrote a micro benchmark to measure the latency w/ and w/o the patch on
kunpeng920 with 96 cores(2 socket, each socket has 2dies, each die has
24 cores, cpu0-cpu47 belong to socket0, cpu48-95 belong to socket1) by
sending IPI to cpu0-cpu95 1000 times from an specified cpu:

#include <linux/module.h>
#include <linux/timekeeping.h>

static void ipi_latency_func(void *val)
{
}

static int __init ipi_latency_init(void)
{

ktime_t stime, etime, delta;
int cpu, i;
int start = smp_processor_id();

stime = ktime_get();
for ( i = 0; i < 1000; i++)
for (cpu = 0; cpu < 96; cpu++)
smp_call_function_single(cpu, ipi_latency_func, NULL, 1);
etime = ktime_get();

delta = ktime_sub(etime, stime);

printk("%s ipi from cpu%d to cpu0-95 delta of 1000times:%lld\n",
__func__, start, delta);

return 0;
}
module_init(ipi_latency_init);

static void ipi_latency_exit(void)
{
}
module_exit(ipi_latency_exit);

MODULE_DESCRIPTION("IPI benchmark");
MODULE_LICENSE("GPL");

do the below 10times:
# taskset -c 0 insmod test.ko
# rmmod test

and the below 10times:
# taskset -c 48 insmod test.ko
# rmmod test

by taskset -c, I can change the source cpu sending IPI.

The result is as below:

vanilla kernel:
[ 103.391684] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122237009
[ 103.537256] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121364329
[ 103.681276] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121420160
[ 103.826254] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122392403
[ 103.970209] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122371262
[ 104.113879] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122041254
[ 104.257444] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121594453
[ 104.402432] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122592556
[ 104.561434] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121601214
[ 104.705561] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121732767

[ 124.592944] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147048939
[ 124.779280] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147467842
[ 124.958162] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146448676
[ 125.129253] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:141537482
[ 125.298848] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147161504
[ 125.471531] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147833787
[ 125.643133] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147438445
[ 125.814530] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146806172
[ 125.989677] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145971002
[ 126.159497] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:147780655

patched kernel:
[ 428.828167] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122195849
[ 428.970822] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122361042
[ 429.111058] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122528494
[ 429.257704] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121155045
[ 429.410186] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121608565
[ 429.570171] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121613673
[ 429.718181] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121593737
[ 429.862615] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:121953875
[ 430.002796] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122102741
[ 430.142741] ipi_latency_init ipi from cpu0 to cpu0-95 delta of 1000times:122005473

[ 516.642812] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145610926
[ 516.817002] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145878266
[ 517.004665] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145602966
[ 517.188758] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145658672
[ 517.372409] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:141329497
[ 517.557313] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146323829
[ 517.733107] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146015196
[ 517.921491] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146439231
[ 518.093129] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:146106916
[ 518.264162] ipi_latency_init ipi from cpu48 to cpu0-95 delta of 1000times:145097868

So there is no much difference between vanilla and patched kernel.

What really makes me worried about my hardware is that IPI sent from the second socket
always shows worse performance than the first socket. This seems to be a problem
worth investigation.

Thanks
Barry

\
 
 \ /
  Last update: 2022-02-20 00:47    [W:0.075 / U:2.276 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site