Re: [Starpu-devel] Task Bench paper accepted

[Samuel Thibault <samuel.thibault@inria.fr>]

Hello,

Elliott Slaughter, le mar. 30 juin 2020 12:58:27 -0700, a ecrit:
> It's been a while so I'm not sure if you remember, but we wrote a benchmark 
> for
> distributed system performance, where we included results for StarPU. I 
> wanted
> to let you know that the paper has been accepted to be published in SC20. 
> That
> means we have about a 2 month window to update results for the camera ready
> paper.

Ah!

> I'm not aware of any outstanding performance issues with the StarPU
> implementation, but I just wanted to close the loop and make sure that you 
> had
> a chance to see the paper one more time and make sure things look ok.

With the current programming model of your benchmark, it is not
unreasonable. However, I had on my TODO list for months now to improve
the programming model used by the benchmark, because it is not actually
making all use of the possibilities of StarPU and of the information
that other runtimes being tested have, notably tags and such, which
depart from the STF programming model. That could make a second StarPU
curve in your graph, to express that it's more the STF programming
model which induces task submission/pruning overhead, than the
runtime overhead. I'll try to get it worked out soon, since that can
dramatically improve the StarPU results.

Thanks for closing the loop :)
Samuel