Hi,
Andi recently added exclusive event group support to tools/perf:
https://lkml.kernel.org/r/[email protected]
and promptly found that they didn't work as specified.
(sorry for the resend, I forgot LKML the first time)
On Thu, Oct 29, 2020 at 05:27:19PM +0100, Peter Zijlstra wrote:
> Hi,
>
> Andi recently added exclusive event group support to tools/perf:
>
> https://lkml.kernel.org/r/[email protected]
>
> and promptly found that they didn't work as specified.
>
> (sorry for the resend, I forgot LKML the first time)
hm, it's too late for me to check ;-) but should I be able to do
this with exclusive event.. running both command at the same time:
$ sudo ./perf stat -e cycles:e -I 1000
# time counts unit events
1.002430650 33,946,849 cycles:e
2.004920725 502,399,986 cycles:e (67.57%)
3.007087631 859,745,048 cycles:e (50.00%)
4.009078254 845,860,723 cycles:e (50.00%)
5.011086104 838,457,275 cycles:e (50.00%)
$ sudo ./perf stat -e cycles:e -I 1000
# time counts unit events
1.001665466 848,973,404 cycles:e (50.01%)
2.003658048 856,505,255 cycles:e (50.00%)
3.005658022 842,737,973 cycles:e (50.00%)
4.007657797 844,800,598 cycles:e (50.00%)
jirka
> hm, it's too late for me to check ;-) but should I be able to do
> this with exclusive event.. running both command at the same time:
Yes. The exclusive part only applies during a given context,
but the two commands are different contexts.
You would only see a difference when in the same context,
and you have multiple groups (or events) that could in theory schedule
in parallel
e.g. something like
perf stat -e '{cycles,cycles},{cycles,cycles}' ...
The main motivation is actually that the "multiple groups" algorithm
in perf doesn't work all that great: it has quite a few cases where it
starves groups or makes the wrong decisions. That is because it is very
difficult (likely NP complete) problem and the kernel takes a lot
of short cuts to avoid spending too much time on it.
With exclusive it will be possible for a tool to generate "perfect groups"
in user space and assume the kernel schedules it dumbly, but at least
without any starvation.
-Andi
On Sun, Nov 01, 2020 at 07:52:38PM -0800, Andi Kleen wrote:
> The main motivation is actually that the "multiple groups" algorithm
> in perf doesn't work all that great: it has quite a few cases where it
> starves groups or makes the wrong decisions. That is because it is very
> difficult (likely NP complete) problem and the kernel takes a lot
> of short cuts to avoid spending too much time on it.
The event scheduling should be starvation free, except in the presence
of pinned events.
If you can show starvation without pinned events, it's a bug.
It will also always do equal or better than exclusive mode wrt PMU
utilization. Again, if it doesn't it's a bug.
Please provide concrete examples for these two cases, or stop spreading
FUD like this.
On Mon, Nov 02, 2020 at 03:16:25PM +0100, Peter Zijlstra wrote:
> On Sun, Nov 01, 2020 at 07:52:38PM -0800, Andi Kleen wrote:
> > The main motivation is actually that the "multiple groups" algorithm
> > in perf doesn't work all that great: it has quite a few cases where it
> > starves groups or makes the wrong decisions. That is because it is very
> > difficult (likely NP complete) problem and the kernel takes a lot
> > of short cuts to avoid spending too much time on it.
>
> The event scheduling should be starvation free, except in the presence
> of pinned events.
>
> If you can show starvation without pinned events, it's a bug.
>
> It will also always do equal or better than exclusive mode wrt PMU
> utilization. Again, if it doesn't it's a bug.
Simple example (I think we've shown that one before):
(on skylake)
$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e instructions,cycles,frontend_retired.latency_ge_2,frontend_retired.latency_ge_16 -a sleep 2
Performance counter stats for 'system wide':
654,514,990 instructions # 0.34 insn per cycle (50.67%)
1,924,297,028 cycles (74.28%)
21,708,935 frontend_retired.latency_ge_2 (75.01%)
1,769,952 frontend_retired.latency_ge_16 (24.99%)
2.002426541 seconds time elapsed
The second frontend_retired should be both getting 50% and the fixed events should be getting
100%. So several events are starved.
Another similar example is trying to schedule the topdown events on Icelake in parallel to other
groups. It works with one extra group, but breaks with two.
(on icelake)
$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e '{slots,topdown-bad-spec,topdown-be-bound,topdown-fe-bound,topdown-retiring},{branches,branches,branches,branches,branches,branches,branches,branches},{branches,branches,branches,branches,branches,branches,branches,branches}' -a sleep 1
Performance counter stats for 'system wide':
71,229,087 slots (60.65%)
5,066,320 topdown-bad-spec # 7.1% bad speculation (60.65%)
35,080,387 topdown-be-bound # 49.2% backend bound (60.65%)
22,769,750 topdown-fe-bound # 32.0% frontend bound (60.65%)
8,336,760 topdown-retiring # 11.7% retiring (60.65%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
424,584 branches (70.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
3,634,075 branches (30.00%)
1.001312511 seconds time elapsed
A tool using exclusive hopefully will be able to do better than this.
-Andi
On Mon, Nov 02, 2020 at 06:41:43PM -0800, Andi Kleen wrote:
> On Mon, Nov 02, 2020 at 03:16:25PM +0100, Peter Zijlstra wrote:
> > On Sun, Nov 01, 2020 at 07:52:38PM -0800, Andi Kleen wrote:
> > > The main motivation is actually that the "multiple groups" algorithm
> > > in perf doesn't work all that great: it has quite a few cases where it
> > > starves groups or makes the wrong decisions. That is because it is very
> > > difficult (likely NP complete) problem and the kernel takes a lot
> > > of short cuts to avoid spending too much time on it.
> >
> > The event scheduling should be starvation free, except in the presence
> > of pinned events.
> >
> > If you can show starvation without pinned events, it's a bug.
> >
> > It will also always do equal or better than exclusive mode wrt PMU
> > utilization. Again, if it doesn't it's a bug.
>
> Simple example (I think we've shown that one before):
>
> (on skylake)
> $ cat /proc/sys/kernel/nmi_watchdog
> 0
> $ perf stat -e instructions,cycles,frontend_retired.latency_ge_2,frontend_retired.latency_ge_16 -a sleep 2
>
> Performance counter stats for 'system wide':
>
> 654,514,990 instructions # 0.34 insn per cycle (50.67%)
> 1,924,297,028 cycles (74.28%)
> 21,708,935 frontend_retired.latency_ge_2 (75.01%)
> 1,769,952 frontend_retired.latency_ge_16 (24.99%)
>
> 2.002426541 seconds time elapsed
>
> The second frontend_retired should be both getting 50% and the fixed events should be getting
> 100%. So several events are starved.
*should* how? Also, nothing is 0% so nothing is getting starved.
> Another similar example is trying to schedule the topdown events on Icelake in parallel to other
> groups. It works with one extra group, but breaks with two.
>
> (on icelake)
> $ cat /proc/sys/kernel/nmi_watchdog
> 0
> $ perf stat -e '{slots,topdown-bad-spec,topdown-be-bound,topdown-fe-bound,topdown-retiring},{branches,branches,branches,branches,branches,branches,branches,branches},{branches,branches,branches,branches,branches,branches,branches,branches}' -a sleep 1
>
> Performance counter stats for 'system wide':
>
> 71,229,087 slots (60.65%)
> 5,066,320 topdown-bad-spec # 7.1% bad speculation (60.65%)
> 35,080,387 topdown-be-bound # 49.2% backend bound (60.65%)
> 22,769,750 topdown-fe-bound # 32.0% frontend bound (60.65%)
> 8,336,760 topdown-retiring # 11.7% retiring (60.65%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 424,584 branches (70.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
> 3,634,075 branches (30.00%)
>
> 1.001312511 seconds time elapsed
>
> A tool using exclusive hopefully will be able to do better than this.
I don't see how, exclusive will always result in equal or worse PMU
utilization, never better.