Re: [Starpu-devel] pheft problem?

[Miguel Palhas <mpalhas@gmail.com>]

Sorry, but i'm no longer working on that project (was my master thesis, finished in september), and i'm now unable to reproduce the error

Since i got no feedback here, i gave up on pheft at the time. Once the implementation was mostly finished i thought about giving it another shot and i was working again, although for some inputs it would crash on random ocasions (but again, only pheft had this problem)

About the hanging up i have no way of telling what was happening, nor what did i change to make it work again

The crashing that came later, i can only tell it happened most often when i was using the weaker of the 2 CUDA devices, and when the amount of simultaenous kernels was higher. So possibly a memory issue?

On Tue, Nov 5, 2013 at 2:02 PM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Hello,

It seems nobody had a look at this issue.

Miguel Palhas, le Tue 13 Aug 2013 09:32:27 +0200, a écrit :

>     When using pheft on my algorithm, i have a problem with a specific task,
>     which hangs the scheduler as soon as it's sampling information has 10
>     executions
>

> This only happens when i set the codelet for that task to use only STARPU_CUDA
> (while it depends on data that is ALWAYS executed on a CPU task)

>     Is this a bug with the pheft scheduler?

Possibly.  I have however just added a testcase
(tests/parallel_tasks/cuda_only), it does finish fine. How are you
specifying CUDA only, through where or through can_execute? (the latter
was buggued for pheft, I have just fixed it).

Could you take a gdb thread apply all bt trace, so we get an idea of
what it is stuck on?

Samuel

--
Cumprimentos

Miguel Palhas