Re: [Starpu-devel] Performance decreasing by adding empty tasks

[Xavier Lacoste <xavier.lacoste@inria.fr>]

Hello again,

I found a workaround, by managing this buffer MySelf, copying it on all CUDA 
devices at the begining of the run and getting the good one in the CUDA 
kernel (I give it an array d_blocktab of ndevices cuda ptr, and choose the 
right one in the kernel with cudaGetDevice).

With this workaround I loose no time adding my global READ-Only buffer.

XL.

Le 17 févr. 2012 à 09:44, Xavier Lacoste a écrit :

> I runned with gprof on a smaller test case : my read only array is 329968 
> integers.
> I can understand that most of the time is spent with data management, but 
> that's all i can say...
> 
> I attached the gprof output.
> 
> <gprof2.out>
> 
> Le 16 févr. 2012 à 16:35, Samuel Thibault a écrit :
> 
>> Xavier Lacoste, le Thu 16 Feb 2012 16:28:54 +0100, a écrit :
>>>     This is not good either, I though that as the buffer is read-only on 
>>> all tasks that uses it, it wouldn't cost much to manage but when I look 
>>> to my Gantt diagram I spend much time "PushingOutput". Why do StarPU have 
>>> to push that read-only buffer ?
>> 
>> Just for the record: it should not ever be doing that for a read-only
>> buffer. Unless something is perhaps keeping the spinlock of the data for
>> a long time for some reason. Maybe using gprof to know where most of the
>> cpu time is spent would help.
>> 
>> Samuel
> 
> _______________________________________________
> Starpu-devel mailing list
> Starpu-devel@lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel