GSoC: MP contention profiling

Robert Luciani rluciani at acm.org
Sat May 16 04:11:02 PDT 2009


Hi all,

As you probably know, I'll be working on making a profiling suite for measuring
MP contention in the kernel. This project will provide a set of tools with which
to provide consistent metrics on kernel contention, and present that data in a
useful (graphical) way. I'm going to start off with porting schedgraph.py so we
can use it to make nice graphs. Currently I'm waiting for feedback from a few
good developers from other BSD projects on whether they had consistent test
methods to gage progress when they worked on MP scalability, and if so, what
tools did they use. It'd be nice to have some points of reference I think.

The suite would most likely be comprised of a set of synthetic tools that I
would create to stress select subsystems such as IPC, VM (mmap/fork) and network
socket performance. If anyone has any suggestions as to what subsystems are
important to them __I'd very much like to hear it__. Perhaps someone thinks that
the tty code is absolutely the most important. :)

Finally I would combine all of these into an easily run suite which in a
consistent and automated fashion produces nice graphs (3d even?) with weighted
averages and other useful output. If there is time over (which I'm sure) I'd
like to begin work on the proc/thread code using lock-free techniques.

I'll be regularly updating the following site with progress:
http://sites.google.com/site/rluciani/soc-09

I won't be on-line on IRC so much the next 2 weeks because I'm in the middle of
moving and am currently in an empty room with only an ethernet cable and a
mattress. June 1st, all 4 of my exams will be done and I'll be set up in my new
apartment. Looking forward to it :)

-- 
Robert Luciani
Chalmers University of Technology, SWE
Department of Computer Science and Engineering





More information about the Kernel mailing list