Parallelize perfparser

[the8472]

Is your feature request related to a problem? Please describe. Opening large perf recordings (e.g. of a test-suite spawning many binaries) takes minutes, most of the time is spent executing a single hotspot-perfparser thread

Describe the solution you'd like Anything that makes it significantly faster, parallelizing the work being a candidate.

Describe alternatives you've considered I have tried lowering the sampling rate and profiling fewer tests at a time, but that merely reduced it from lunch-break to coffee-break time.

Additional context Tested on hotspot 1.3.0

[milianw]

hey @the8472 - this request is super unhelpful. I would also like to parallelize it, but the problem is simply not easily amendable to being parallelized. sampling events need to be handled in order, the best we could do is potentially parallelize the unwinding for separate processes. patches welcome I guess, but again - not an easy task.

[the8472]

this request is super unhelpful.

I'll see if I can get a flamegraph of the parsing that's slow for me. Would that help?

sampling events need to be handled in order

I don't know anything about the perf data format. Is it not separable into large chunks that can be processed independently and then joined? I had that impression because it occasionally prints the message that some chunks were lost and it's able to recover from that.

the best we could do is potentially parallelize the unwinding for separate processes

That may help in at least one of my cases where I was profiling a benchmark suite which spawned ~10k short-lived, separately compiled processes.

[the8472]

Top-down and bottom-up flamegraphs of hotspot+perf-parser loading a 9GB perf data which takes several minutes.

[milianw]

hey @the8472, thanks for the flamegraphs, but as-is this is still totally unactionable to me. The flamegraphs only show that dwarf resolution of inline frames is slow, as well as repeated mmap/munmap on your system. The latter is surprising, but I'm unsure I can do anything about it. The former is less surprising, as DWARF inline frame resolving is generally a slow thing. If that's needed a ton, then it's simply slow. Why it is so slow for your case compared to other scenarios I looked at so far, I cannot say. Without a way for me to reproduce this issue, I cannot look into it.

If you want me to do anything about this, you will need to either document how you record the perf file such that I can reproduce the issue. Alternatively, you will need to upload the perf.data file as well as all the binaries and libraries and debug files it references - this can easily become a very large tarball requiring multiple gigabytes of storage space. Upload it to a cloud downloader of your choice and share a link here. Note that this obviously only is an option if the binaries included are open source, if that includes proprietary code there's nothing I can do about this task.

[the8472]

This will require beefy hardware or patience:

1. clone https://github.com/rust-lang/rust
2. install dependencies
3. setup a build config. setting download-ci-llvm=true is recommended to reduce build times, but not necessary.
4. additionally set debuginfo-level = 2 in the [rust] section to build with full debuginfo
5. run ./x test --stage 1 ui to get things built
6. run perf record --call-graph dwarf -F 97 -e instructions,cycles ./x test --stage 1 --force-rerun ui to profile the UI testsuite
7. open the profile in hotspot
8. observe slow parser

[milianw]

thanks, I'll see when I can find the time to replicate that environment. But for now, note already that you should change your perf invocation to leverage leader sampling. I.e. right now, you program the PMU to sample both, instructions and cycles independently at ~97 samples per second. This is probably not what you want - instead, write it like this:

perf record --call-graph dwarf -F 97 -e "{cycles,instructions}:S" ...

this will sample on cycles and then whenever that happens also record the instruction count too.

[milianw]

Hm I cannot compile the rust compiler, it seems to not honor PATH but instead just looks only in the first entry for ar instead of continuing onwards?

$ ./x test --stage 1 src/test/ui
Building rustbuild
    Finished dev [unoptimized] target(s) in 0.03s
thread 'main' panicked at '

couldn't find required command: "/home/milian/.bin/ar"

', sanity.rs:59:13
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
Build completed unsuccessfully in 0:00:00
$ which ar
/usr/bin/ar
$ echo $PATH
/home/milian/.bin:/home/milian/.local/bin:/home/milian/projects/compiled/other/bin:/home/milian/.bin/kf5:/home/milian/projects/compiled/kf5-dbg/bin:/home/milian/projects/compiled/other/bin:/home/milian/projects/compiled/kf5/bin:/usr/local/sbin:/usr/local/bin:/usr/bin:/usr/lib/jvm/default/bin:/usr/bin/site_perl:/usr/bin/vendor_perl:/usr/bin/core_perl

[the8472]

I think the build tools try to infer the ar path based on the c compiler path. If the default logic doesn't work you can explicitly set its path in config.toml

[GitMensch]

@the8472: Would you mind to record as per the suggestion above and recheck the loading with a current appimage?

[the8472]

Which suggestion are you referring to?

[the8472]

-e "{cycles,instructions}:S"

Using this leads to tiny profiles (less than a megabyte) instead of gigabytes and opening such a profile crashes hotspot.

[GitMensch]

I can't reproduce that with some simple examples:

perf record --call-graph dwarf -F 97 -e "{cycles,instructions}:S" true
perf record --call-graph dwarf -F 97 true

both open fine (same for a bigger application).

Can you please share the perf file that crashes hotspot for you (if not already tested, please check the latest Appimage) and how you recorded that.

[the8472]

I'm rerunning my steps from https://github.com/KDAB/hotspot/issues/394#issuecomment-1312800881 only modifying the record command:

perf record --call-graph dwarf -F97 -e "{cycles,instructions}:S" ./x test ui --stage 1 --force-rerun

perf.data.gz

It crashes both hotspot-v1.4.1-263-ga8d1440-x86_64.AppImage and hotspot-v1.3.0-277-g2bcd36d-x86_64.AppImage with a segfault

/tmp/.mount_hotspoWNLRPL/AppRun.wrapped: line 6: 2288130 Segmentation fault (core dumped) LD_LIBRARY_PATH="$d/usr/lib":$LD_LIBRARY_PATH "$bin/hotspot" "$@"

[milianw]

I think the crashes should be resolved in newer hotspot. But the fact that the data files are tiny sounds like a kernel bug - is that still the case now?

[GitMensch]

Rechecked: appimage from October has that fault as above after loading "long", using the appimage from today hotspot directly opens, no chrash any longer.

For the original case we likely still need a more reproducible step :-/

[the8472]

For the original case we likely still need a more reproducible step :-/

Can you explain what is not reproducible about https://github.com/KDAB/hotspot/issues/394#issuecomment-1312800881 ?

[milianw]

It's just hard to setup, but I think it should be enough. I just need to find the time to replicate your setup, which is a hurdle for me. Last time I tried I failed, and now I just need to try again with your suggestion on how to workaround that issue.

[the8472]

Would Dockerfile containing all the steps to reproduce it help? It'll be a bit tricky to also get the GUI to run in docker but I think it should be possible.

[milianw]

A docker file would help, I can just access it from the outside then using hotspot to analyze the data. I.e. the docker image just needs rust + perf and no UI at all I think.

[the8472]

FROM archlinux

RUN pacman -Sy && pacman -S --noconfirm git perf base-devel

WORKDIR /opt
RUN git clone --depth 1 https://github.com/rust-lang/rust --branch master --single-branch rust

WORKDIR /opt/rust
RUN cp src/bootstrap/defaults/config.compiler.toml config.toml

CMD ./x test ui --stage 1 -- "does not exist" && perf record -m 128 -F59 --call-graph dwarf -e cycles:u ./x test ui --stage 1

podman build  -t perfparser-test .
podman run -it --privileged perfparser-test

The privileged flag is required to run perf record.

[milianw]

Alright, I looked at this a bit now. It might theoretically help for this specific problem if we could parallelize the perfparser analysis on the pid level, since there are so many of them in the above record.

Generally though, there is a ton of inherent overhead that I would like to see removed/reduced, but I fear elfutils is not prepared for that - most notably we load the same libraries and seem to spend the most time during unwinding on finding CFI data and building some elfutils-specific cache, only to then remove that again and rebuilding it for the following process in quick succession. I.e. the problem really is that there are thousands of short running processes here, which is apparently the worst case for analysis.