search

plasma-umass / coz

Coz: Causal Profiling
Other
4.08k stars 159 forks source link

Fail when a --fixed-line is given and it is not found #145

Closed amedeedaboville closed 4 years ago

amedeedaboville commented 4 years ago

Hi, thanks for making coz!

I've been using it a bit on toy programs, with coz-rs.

This PR: I think the run should fail if we don't find the fixed line we're told to experiment on. Do you agree?

Also, I'd like to write+submit PRs for these, what do you think of these ergonomic improvements?

I have this locally and it has helped me a lot because often with Rust (and maybe its closures or all the macro usage), not all lines end up showing up. Knowing which lines are accessible helps a lot.

For example, one might want to compare several different locations in a program at once, instead of in separate runs.

A natural ergonomic follow-up:

This might be useful because in many programs you don't really expect to speedup most functions by more than 40%, so it might speed up runs by needing to collect less data / instead collecting more data to get more accuracy on the range you care about.

An even further solution might even allow passing an extra step value (eg --fixed-speedup 0-40:2 = 0 to 40 by steps of 2) if someone wants control over the granularity of the speedups.

Let me know what you think and I'll get to submitting them! I might also drop a couple other small PRs for papercuts I've encountered so far.

ccurtsinger commented 4 years ago

Thanks for the fix; not many people are using fixed-line experiments so this is helpful from a usability standpoint.

It would be great to have support for speedup ranges. For now I think it should choose randomly using the existing setup, but I see your issue (#146) related to this topic. I'll need to think it over before I can give a definite answer. It's important that speedup amounts include data for zero speedup even if its not in the specified range, since that establishes the baseline.