Hello! Week 5 of profiler work is over. as a reminder – what I’ve been doing is building a new sampling CPU profiler for Ruby! It’s called rbspy and it’s at https://github.com/rbspy/rbspy.
In just-this-second news – someone tweeted at me just now that they used rbspy and it helped them find an unexpected hot spot in their program, which made me super happy!!
The main 2 exciting things that happened last week were:
- Mac support is done! Supporting Mac is kind of an interesting thing because in my mind rbspy is a production profiler (for figuring out why your production server Ruby code is slow), and Macs are basically all laptops. Nonetheless! Supporting Mac is cool, people are using the Mac version (to make perf improvements to CocoaPods, which is a Mac Ruby program) and it’s working and I’m happy about that.
- I’ve been doing a lot of experiments with memory profiling. Basically I’m curious about what memory profiling tools I can build that work from outside of the Ruby process (just give the tool a PID and it tells you about the internals of your Ruby program!)
memory profiling experiment: tracking line numbers of every allocation
Here’s an experiment I worked on today! I wanted to be able to answer the question – which functions are allocating memory right now?
So I wrote a small program that tracks every Ruby memory allocation, collects the function, filename, and line number that that allocation happens at, and then aggregates the results! This happens live, and should work on basically Ruby program with symbols.
here’s the example output for a run of
rubocop I did:
635 wrap_with_sgr /home/bork/.rbenv/versions/2.4.0/lib/ruby/gems/2.4.0/gems/rainbow-3.0.0/lib/rainbow/string_utils.rb 13 898 tok /home/bork/.rbenv/versions/2.4.0/lib/ruby/gems/2.4.0/gems/parser-184.108.40.206/lib/parser/lexer.rb 21731 931 visit_descendants /home/bork/.rbenv/versions/2.4.0/lib/ruby/gems/2.4.0/gems/rubocop-0.52.1/lib/rubocop/ast/node.rb 552 1262 source /home/bork/.rbenv/versions/2.4.0/lib/ruby/gems/2.4.0/gems/parser-220.127.116.11/lib/parser/source/range.rb 129
This says that there were 1262 allocations in the
source function in
range.rb on line 129. Let’s
look at line 129, to see what that looks like:
128| def source 129| @source_buffer.slice(self.begin_pos...self.end_pos) <---- line 129 130| end
It’s not immediately obvious what the allocation is here, but it turns out that it’s
self.begin_pos...self.end_pos – that allocates a new
Range object. So that’s kind of
This particular example doesn’t let us make any actual performance improvements to rubocop (1000 allocations over 10 seconds is actually not a lot at all), but it’s interesting to be able to see which lines are allocating the most.
this week: debian packaging, more experiments
One goal for this week is to make a Debian PPA for rbspy – I think that’ll be really helpful for Linux users (who are the main target audience I have in mind). There’s also 1 bug with C stack frames that I’ve been meaning to fix for a while. I feel like I’m mostly done (?) with core rbspy functionality and that it’s going to become time to focus more on UI and other non-CPU-profiler experiments.
One UI concern that I have right now is – flamegraphs are a nice way to present profiler data, but they don’t work well for recursive programs. So I might try to put together another visualization style, like this style that pprof generates: (image taken from the golang blog).
I think for the rest of the sabbatical I’ll split my time between stabilizing rbspy / adding UX improvements to it and working on new experiments in memory profiling / other ways to observe what Ruby programs are doing.
If you have ideas for experiments in the future of Ruby profiling, I’d love to hear them!
want to help?
A few people have said they’d love to help with rbspy. In case that’s you, a very useful thing you can do is try it out and report your successes/issues!