With a slight delay caused by going to lovely lovely Istanbul for the
LREC conference where I presented a poster, I am back to work on the
Google Summer of Code project. By the way, this year’s logo and swag
looks a lot nicer than last year’s, thank you Google!
The backbone of my GSoC consists of putting together a continuous benchmark platform. I took a good look at vbench and spent an evening hacking Wes’s benchmarks suite config into something that will run on my machine. These are the key points I got from this experience.
- vbench is, at least for the moment, very specific to [Wes’ and Pandas’ needs]. This is also because there weren’t so many other users that could have brought contributions.
- Even though it has support for some configuration and automation, vbench seems largely suited for running on a local machine. Specifically, it is NOT designed to run continuously but in one-off runs, going back in git history and getting the last commit for each day, and running the benchmark with it. Of course, it is trivial to patch it into getting just one commit.
- The code-as-strings approach is not ideal. The first thought is
that it should be replaced with reading
.pyfiles into strings, but there are two issues with this:
- One benchmark file can have a lot of setup code and several key
lines that need to actually be benched. This can be fixed using
convensions (ie. setup functions and
bench_*functions) in the spirit of testing suites, or using decorators.
- I would like to be able to run bench files as python scripts, but the vbench import system breaks this. This can be fixed by hijacking the imports when reading the file.
- One benchmark file can have a lot of setup code and several key lines that need to actually be benched. This can be fixed using convensions (ie. setup functions and
Our project has different dynamics than Pandas, so it’s important that the published results run on an independent machine, but it would be great if an individual developer can run the benchmark himself while coding but before pushing his changes upstream. Of course, his numbers would only be comparable to the numbers he gets on his own machine before his changes, but a developer shouldn’t wait for the daily benchmark for knowing if he made an improvement.
- Memory usage benchmarking
- Python scripts as benchmarks, with a simple but efficient Benchmark object hierarchy
What’s missing is:
- A system to remember previous results and compare them, similar to vbench’s database
- The ability to bench only an area of the code without rerunning the setup. (Not really sure whether vbench’s way is actually better)
At a first glance, it seems that a very good system can be obtained by
combining these two excellent projects (or rather, improving vbench with
perf.py). While I continue exploring this, I would like
to hear feedback from people who had to do with similar issues. As for
the GSoC timeline, I plan to join forces with Immanuel and design a
solid benchmark suite for the linear models over the next 2 weeks.