Re: [Starpu-devel] memory overhead

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Hello,

Xavier Lacoste, le Fri 06 Dec 2013 14:53:10 +0100, a écrit :
> Have you go an idea of what could explain this overhead (+8% compared to 
> PaStiX
> peak but much more if we don't take into account common data structures, 
> which
> may be a more interesting metric....) and the huge number of calls to 
> malloc().

Well, we have never spent time on optimizing allocations, so it is
generally not really surprising: we would have a few mallocs for each
task and each data transfer, for instance.

What is particular in your case, however, is that you have a performance
model with a lot of different sizes, which are recorded separately.
That's also one of the sources of overhead at the start of the
application, just to load the multiple-MB performance model file. It
would probably be useful to have a look at using a regression-based
performance model to avoid storing so many measurements.

Samuel