I've created instructions on building v8 and running a molecular
dynamics engine we are working on with various profiling flags turned
on:

  https://gist.github.com/2289269

I'm working to understand the serious performance regression in Chrome
I reported here:

Huge performance regression/variation in Chrome after changing window
size and reloading (using typed arrays and SVG)
http://code.google.com/p/chromium/issues/detail?id=121196

I do not yet know if the data I am collecting also shows the kind of
slowdowns we are seeing in Chrome -- but is it already clear that
something is causing the main integration function to be deoptimized
over and over:

  makeIntegrator.integrate()

See the profile.log here: https://raw.github.com/gist/2289269/profiling.log

I've been reading Florian Loitsch's series on profiling and optimizing
code for V8 starting with this post:

  http://floitsch.blogspot.com/2012/03/optimizing-for-v8-introduction.html

But I'm still trying to understand what is causing the
deoptimization ... the deoptimization of makeIntegrator.integrate()
makes the code run much slower than it needs to.

I'm getting a series of reports like these:

    **** DEOPT: makeIntegrator.integrate at bailout #436, address 0x0,
frame size 1004
                ;;; @2986: check-maps.
    [deoptimizing: begin 0x2b41155d makeIntegrator.integrate @436]
      translating makeIntegrator.integrate => node=1794, height=112

    ...

    **** DEOPT: makeIntegrator.integrate at bailout #213, address 0x0,
frame size 984
                ;;; @1546: deoptimize.
    [deoptimizing: begin 0x2b41155d makeIntegrator.integrate @213]
      translating makeIntegrator.integrate => node=853, height=132

Is there more documentation for understanding results like:

  ;;; @2986: check-maps

and

  ;;; @1546: deoptimize

I do not yet have good profile data to make any definitive
interpretation of the following results .. but I am normally seeing
the model itself running at 550-600 model-steps/s (Mac OS X 10.6.8,
2.66 GHz Intel Core i7, w/8GB memory).

When Chrome is running slowly the rate drops to 120-150 ... however
during one test on Windows 7 (on the same machine) I got a measurement
of 1724 model-steps/s.

It would be nice to get that performance more regularly.

FYI: you can see the model run with a visualization of molecular
dynamics here:

http://concord-consortium.github.com/lab/examples/simple-atoms-model/simple-atoms-model.html

-- 
v8-users mailing list
[email protected]
http://groups.google.com/group/v8-users

Reply via email to