60x speed-up of Linux “perf”

Re:60x speed-up of Linux “perf” by vlovich123 on Thursday September 09, 2021 @10:17AM (#28469942)

I haven't blogged about this yet, but we saw a 1000x fold speed-up doing several things around symbolication. The more optimal approach we found was to use the gimli crate[1] directly & carefully optimize it to read in the data structures for the executable(s) you are symbolicating upfront & then issuing in-process queries. They also have a drop-in replacement of addr2line that outperforms it (both in symbolication speed & memory usage).

I am curious about how that compares with libbfd since that wasn't under consideration for us as it uses GPLv3.

[1] https://github.com/gimli-rs/gimli

[2] https://github.com/gimli-rs/addr2line

[ Parent | on HN ]

Re:60x speed-up of Linux “perf” by tonyg on Thursday September 09, 2021 @10:35AM (#28470140)

Steinar Gunderson (who suggested [1] the talk-through-a-pipe approach) just now compared [2] the `libbfd` variant to the pipe-to-`addr2line` version, and found them to take similar amounts of time.

This agrees with what I saw in `top` while testing: with the patch, I see `perf` using ~95% CPU, with `addr2line` using the remaining ~5%.

So speeding up `addr2line` probably wouldn't result in very much of an overall improvement for this workload.

[1]: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=911815#28 [2]: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=911815#38

[ Parent | on HN ]

Re:60x speed-up of Linux “perf” by vlovich123 on Thursday September 09, 2021 @11:07AM (#28470501)

The piping is a good chunk of the gain but the gimli library is still much faster than addr2line & uses less memory if used correctly. Since libbfd shows similar results, I'm thinking there might be a speed bump from switching to use it (some care has to be taken to preserve 0-copy semantics across FFI).

[ Parent | on HN ]

Re:60x speed-up of Linux “perf” by tonyg on Thursday September 09, 2021 @11:10AM (#28470545)

It'd definitely be interesting to find out!

[ Parent | on HN ]

Re:60x speed-up of Linux “perf” by vient on Thursday September 09, 2021 @11:20AM (#28470661)

Looks interesting. I've encountered some annoying bugs with latest addr2line when trying to get symbols for addresses, wonder how gimli performs in that sense. Also, it seems that addr2line 2.37 has some performance regressions compared to 2.36.1 which I didn't look at yet.

[ Parent | on HN ]

Re:60x speed-up of Linux “perf” by vlovich123 on Thursday September 09, 2021 @12:00PM (#28471149)

In my experience gimli also symbolicated things addr2line couldn't.

[ Parent | on HN ]