Re: [Starpu-devel] Failed assert in copy_driver.c

[Xavier Lacoste <xl64100@gmail.com>]

Le 27 oct. 2014 à 17:32, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
écrit :

> Samuel Thibault, le Mon 27 Oct 2014 17:29:22 +0100, a écrit :
>> Xavier Lacoste, le Fri 24 Oct 2014 02:37:26 +0200, a écrit :
>>> I wen’t back to r13630 because timing are better with this revision when 
>>> GPUs are used.
>> 
>> Ah?  It'd be good to investigate.  Which scheduler and matrix are you 
>> using?
> 
> Ah, sorry, I guess these are lws and the various matrices shown on your
> picture.
Yes, I forgot I uploaded the picture, thats right : "lws" on the revision 
0346c41 of PaStiX.

Since that commit I improved the timing a bit by giving the right priorities 
(the ones of PaStiX static scheduler) to each tasks (I didn't filled this 
field for each). I also gave MAX_INT priority to communication tasks to force 
the consumption of the received data (and also send data as soon as they are 
ready). Indeed the run failed on some matrix due to lack of memory and we 
supposed that the received data are not used as soon as they are received. I 
still have to do a memory study to confirm that. This may also explain why 
the fan-out version gives good performance while we didn't expected it would.

Also, I still have to fix a numerical issue with the LDLt version on GPU 
(Matrix Serena / pml_DF) , so better avoid these matrices...

> 
> Samuel
> _______________________________________________
> Starpu-devel mailing list
> Starpu-devel@lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail