perf_events

[jbboehr]

Never start by creating a pull request, always open an issue to explain what you want to do before starting developing something.

I've been messing around with this extension recently and it's been really great! I am looking into using it for profiling my application over time, but I am a little concerned the metrics (wt, ct) might be a little noisy.

perf seems to have the ability to measure instructions and cycles, which should be fairly stable, as long as the hardware and php compiler flags don't change. Would there be any interest in adding these metrics?

perf, the program, seems a little tricky to use since it is coupled to the linux kernel version. I don't know if perf_event_open suffers from this as well, or not.

HHVM seems to use it somehow: https://github.com/facebook/hhvm/blob/master/hphp/util/perf-event.cpp
Manpage for perf_event_open: https://man7.org/linux/man-pages/man2/perf_event_open.2.html

I wouldn't mind looking into it a bit if you are interested.

[jbboehr]

Another fairly stable metric should be "number of executed zend opcodes". It looks like there is this:

zuo Zend Engine user opcode count Number of included userland opcodes (sum of all userland file/function/method opcodes).

However that seems to only show how many are loaded, not actually executed.

[NoiseByNorthwest]

I don't know if perf_event_open suffers from this as well, or not.

I haven't read everything here https://man7.org/linux/man-pages/man2/perf_event_open.2.html, but the worst case seems to be Linux 4.3+

However that seems to only show how many are loaded, not actually executed.

Exactly, and there is no such stat collected by Zend Engine.

I wouldn't mind looking into it a bit if you are interested.

Yes I could be interested, but could you first:

• explain to me in which circumstance the noise of wt/ct is an issue ? (Real question, I'm curious)
• confirm me that getting current CPU cycle count via perf has a reasonably low overhead and could be done thousands or millions of times per second
• confirm me that the result will really be that stable. For instance will the number of context switches have an influence on CPU cycle count ? Or will the perf instrumentation itself generate noise ?
• confirm me that the cycle count is for the current process whatever the CPUs used, in other word that it would not require to stick to a given CPU

[jbboehr]

tl;dr It could be useful for profiling, but I think my needs might be better served by e.g. phpbench. I'm still planning on using spx for tracing, though ❤️

---

Exactly, and there is no such stat collected by Zend Engine.

After I posted this, I poked around in php-src for it and couldn't find it either. Closest was tick_count, but that requires ticks to be enabled and is a big performance hit, IIRC.

explain to me in which circumstance the noise of wt/ct is an issue ? (Real question, I'm curious)

I'm trying to track performance over time in a potentially noisy CI environment. I realized later that my use case might be better served by e.g. phpbench. Although they don't seem to be able to track custom metrics yet...

confirm me that getting current CPU cycle count via perf has a reasonably low overhead and could be done thousands or millions of times per second

Honestly, this may be a bit above my pay grade, but I suspect that using it in sampled mode should be fairly low overhead, as the kernel appears to write the data into a shared memory buffer (I believe the HHVM link above uses this). It appears a mutex and a signal handler will be necessary in this case.

I experimented a bit with exposing the info without sampling to userland. In this case, it involves one read(2) call + optionally(?) two ioctl if you want to pause the counters while processing that data + processing the buffer.

confirm me that the cycle count is for the current process whatever the CPUs used, in other word that it would not require to stick to a given CPU

I'm not sure, maybe this part from the manpage would clear things up a bit:

  pid == 0 and cpu == -1
         This measures the calling process/thread on any CPU.

confirm me that the result will really be that stable. For instance will the number of context switches have an influence on CPU cycle count ? Or will the perf instrumentation itself generate noise ?

I'm not sure about the context switches. I would imagine context switches shouldn't affect the number of executed instructions that much, but I don't know. There will be some overhead from reading the data, of course. The example pauses the counters with ioctl(fd, PERF_EVENT_IOC_DISABLE); while processing the data, which will still, of course, affect the wall time, but would reduce the effect on the perf counters themselves.

---

Just to reiterate, this is a linux syscall so it would have to be behind a configure flag (no BSD, OSX, Windows, etc).

[jbboehr]

Looks like perf_event_open is unavailable on Github Actions, might not be worth the trouble: actions/runner-images#4974

(The CI system I was planning to run it on was not Github Actions, though)

[NoiseByNorthwest]

tl;dr It could be useful for profiling, but I think my needs might be better served by e.g. phpbench. I'm still planning on using spx for tracing, though ❤️

Yes it is not the best tool for performance testing.
In conclusion, can we say that we haven't yet identified enough added value to add the support of CPU cycle count to SPX ?

[jbboehr]

I think it could be valuable, but not being able to run it in GHA is a big pain point.

It sounds like you aren't convinced they would be useful enough to outweigh the added potential maintenance, which is totally okay, of course. Thank you for entertaining the idea.

For now, if I find the time and motivation, I think I will play around with my userspace bindings, perhaps I'll find some workarounds.

[NoiseByNorthwest]

It sounds like you aren't convinced they would be useful enough to outweigh the added potential maintenance, which is totally okay, of course. Thank you for entertaining the idea.

I'm not totally closed to the idea, it's just, as you said, a matter of trade-off between all costs and the added value.

I've also seen your repository about the user space binding https://github.com/jbboehr/php-perf. If, in the future, I change my mind, it could help a lot with adding this metric to SPX.