Re: [Starpu-devel] Performance with SOCL on multiple devices

[Malcolm Roberts <malcolm.i.w.roberts@gmail.com>]

Hello.

I've spent some more time getting things to run in parallel, and I've 
had some interesting results.  In particular, I can now launch waves of 
DGMacroCellInterface kernels in parallel.  However, this seem to do 
nothing but slow things down.  I used nvvp to visualize what the 
scheduling looked like, and it seems that launching things in parallel 
(with lots of OpenCL events in each kernel's wait list) adds a lot of 
overhead.  The inter-kernel time was around 50 times larger than when I 
was launching kernels in sequence.  I looked at what this looks like 
under SOCL as well, and it seems that this is the same problem; the 
overhead involved in launching kernels delays the start of execution, 
and it's this which is causing the performance degradation.

I'm not sure what the best strategy is right now, but I think that I'll 
spend some time on improving the individual kernels while I think about 
what to do next.

Thanks for your help!

Best,

~Malcolm

On 06/01/2016 17:16, Samuel Thibault wrote:
> Samuel Thibault, on Wed 06 Jan 2016 17:03:42 +0100, wrote:
> > With these changes, I do see kernels running on various GPUs. It does
> > not seem faster, but that is probably due to data transfers.
> And also serialized kernels
>
> > You will probably want to use FxT to dump traces and read them with
> > Vite.
> It notably seems that the DGMacroCellInterface kernels are completely
> serialized, thus awfully bad performance due to no parallelism and
> data transfers :) DGFlux does get parallelized, on the other hand. The
> duration is however quite tiny (60µs), so it's really not efficient
> compared with the runtime overhead.
>
> Samuel

--
http://malcolmiwroberts.com