Publish perf script output somewhere, somehow.

[Mark-Simulacrum]

No description provided.

[Mark-Simulacrum]

@alexcrichton How hard would it be to set up an S3 bucket or similar to which these could be uploaded? They look like about ~500 MB of information per commit, which compresses down to a few MB -- so we shouldn't use too much storage, but the github solution will no longer work well.

Realistically, it's probably time to start thinking about moving all of our data for perf.rlo to either a DB or at least an S3 bucket which we can sync to the server(s).

[alexcrichton]

Oh shouldn't be hard at all, although I think we'd probably want to still have the data expire after 90 days or so. Lemme know if you'd like to do that and I can set up a bucket and give you some keys.

[Mark-Simulacrum]

@nikomatsakis Are you okay with these files going away after 90 days? If so, then let's get the S3 bucket up and running so we can put the script files (and maybe some other metadata in there).

[nikomatsakis]

@Mark-Simulacrum seems ok for now

[Mark-Simulacrum]

@alexcrichton Do you need anything from me to setup the bucket? If not, then please do so and I'll work to get the relevant details implemented in perf's collector.

[alexcrichton]

I've created a bucket where a sample file is https://s3-us-west-1.amazonaws.com/rust-lang-perf/log-1.txt, and the access keys should be in 1password

[Mark-Simulacrum]

So it looks like the next step here is to figure out a way to get both the stats from perf stat and the perf script output from the same run. So far I've not been able to find how to do this...

[Mark-Simulacrum]

@nikomatsakis This is very beta (and seems to not work with perf focus, though the data... seems to be there? I'm not sure what's wrong), but initial files are up: https://s3-us-west-1.amazonaws.com/rust-lang-perf/4be5d360cd8a9338d1408b3998874313240db395/clap-rs-clean-check and wget https://s3-us-west-1.amazonaws.com/rust-lang-perf/4be5d360cd8a9338d1408b3998874313240db395/clap-rs-nll-check for example. Note that these aren't yet being compressed and aren't linked from anywhere on the site, though the URL is relatively straightforward to construct: grab the commit hash you're interested in, then the benchmark name (so clap-rs in this case, this'll be the directory), then the state name (see table here: https://github.com/rust-lang-nursery/rustc-perf/blob/master/collector/src/lib.rs#L132, replacing : with _ and with -), then the suffix is -check, -opt, or nothing for debug builds.

Note that these URLs may change, we may stop uploading artifacts, there may be other issues that I'm as yet unaware of. Basically: super-unstable, don't depend on these existing.

[Mark-Simulacrum]

I've reverted the patch for the time being; it turned out to not work well timing wise, as well as (apparently) causing the OOM killer to kill perf. I'm also not happy with how slow perf record/perf script are; they appear to add ~2-3 hours onto each build which is not acceptable long-term.

There's also a problem with the current "solution" to collecting perf.data files with perf record -- we lose accurate event counts for instructions and the other events we collect. This may be fine (to an extent) but I'd like to do more tests to determine whether that loss of accuracy is something we can afford.

@nikomatsakis Is it possible that there's some small benchmarks (probably not full crates rather the specific test-cases) that we could run the less-accurate but more-dataful collector for? That way we can do that in the meantime and at least have something...

[Mark-Simulacrum]

I'm going to give this a close for now since we're moving towards using self-profiling data rather than perf script, but we may want to revisit this in the future.