[Swift-devel] Re: waitpid mini-benchmark

[Michael Wilde]

Agreed. That was the purpose of measuring it. Probably should make some easy way to pass command line options to the worker.

I think we felt all along that the current logic was OK for workersPerNode <16 or so.

The concern was only related to running workersPerNode much higher, like O(100s), which would only be done for experiments (what we've been calling "over-clocking the nodes" with sleep jobs).  Ie, to use one rack of the BG/P to simulate the client load of 40 racks.

So I agree - this task is off the list.

- Mike


----- Original Message -----
> That's 0.3ms for 100 jobs? If yes, I don't think we should worry about
> it beyond making sure that we don't poll very often (e.g. more than 10
> times/s).
> 
> On Tue, 2011-01-11 at 15:34 -0600, Justin M Wozniak wrote:
> > Just posted the results of a Perl waitpid() mini-benchmark:
> >
> > http://www.ci.uchicago.edu/wiki/bin/view/SWFT/MiniBenchmarks
> >

-- 
Michael Wilde
Computation Institute, University of Chicago
Mathematics and Computer Science Division
Argonne National Laboratory