OpenBSD Journal

g2k19 Hackathon Report: back to dee trace

Contributed by rueda on from the In case of fire, open trace dept.

Martin Pieuchot (mpi@) wrote in with a report from g2k19:

I was particularly excited to attend this year's general hackathon. I arrived without expectation. My goal was to have a fresh look at the kernel and discuss pending issues in a new way. Since I haven't really hacked on OpenBSD during the last year, I wasn't sure what to do next.

So I resumed hacking on a dtrace(1) or bpftrace(8) compatible tool to be able to answer the question. After some exchanges with many developers, I settled on a design and started writing code. Hackathons are to start projects, right? After a couple of days tobhe@ decided to join the effort and started working on a parser for the tool. Then a couple of days later, I had my answer:

Flame Graph

This shows which codepaths are consuming CPU time in the kernel while building libc on a 2 CPU machine. It is now obvious which subsystem needs to be pushed away from the KERNEL_LOCK() in order to reduce contention during builds.

During the week many developers argued that the SCHED_LOCK() was too contended. I don't know which use case triggers such contention, but I gave a shot at splitting the lock. Sadly the change exposed a lock ordering problem, so it has been reverted. So today I enjoyed splitting the SCHED_LOCK() a bit further. I'm not sure if these changes are necessary for now. But I believe that it is more important to explain and show what is possible to do first.

That's also a reason why I spent so much time during the hackathon talking about the possible next steps to push the KERNEL_LOCK() down. I'm happy to see the number of developers working on parallelism grow. I also tried to explain the best I could the non-design of the NET_LOCK() and what could be do to improve parallelism in the Network Stack.

It was a really fun hackathon. I was really happy to see old and new faces. I am so thankful to all those clever people for sharing their time with me.

Thanks a lot to mortimer@ and the OpenBSD foundation for making that hackathon possible!

Thanks very much for the report Martin!

(Comments are closed)


Comments
  1. By Philipp (pb) on

    This is very good news. 1000x for dwelling into this!

Credits

Copyright © - Daniel Hartmeier. All rights reserved. Articles and comments are copyright their respective authors, submission implies license to publish on this web site. Contents of the archive prior to as well as images and HTML templates were copied from the fabulous original deadly.org with Jose's and Jim's kind permission. This journal runs as CGI with httpd(8) on OpenBSD, the source code is BSD licensed. undeadly \Un*dead"ly\, a. Not subject to death; immortal. [Obs.]