Re: [Starpu-devel] [trace interpretation]

[Maxim Abalenkov <maxim.abalenkov@gmail.com>]

Dear all,

I came to a conclusion that my low performance might be caused by very fine task granularity that I have created and too many data transfers between tasks. I’m on a mission to merge some of the tasks together. I would like to ask a question: I know subtasks are not supported in StarPU, but would it be possible to dedicate a certain subset of tasks to perform some computation and let other threads "jump over” this computation and get busy with the subsequent work? For example, if I had 20 threads and 3 algorithmic parts, could I dedicate 4 threads to work on part 1 and let 16 threads proceed with parts 2 and 3? Thank you and have a good evening!

—

Best wishes,

Maxim

Maxim Abalenkov \\ maxim.abalenkov@gmail.com
+44 7 486 486 505 \\ http://mabalenk.gitlab.io

On 24 Jul 2018, at 16:44, Samuel Thibault <samuel.thibault@inria.fr> wrote:

Maxim Abalenkov, le mar. 24 juil. 2018 15:41:10 +0100, a ecrit:

Do I understand correctly, that the green coloured bars are
good—threads are working, red are bad—threads are waiting,

Yes. Double-click on the to get more info.

white arrows are the task dependencies?

They are rather data transfers. Which happen to match task dependencies.

We should probably write down these in some guide in the documentation.

Any hints on how to improve performance in this particular case would be very welcome.

Well, without knowing the algorithm it's quite hard :)

It can however at least be seen at the bottom that task submission is
still happening during the execution, which is probably not what you
want, you probably have undesired synchronization steps in your main()
function.

Samuel