Re: [Starpu-devel] Questions about StarPU worker recongnization

[Fangli Pi <hpcfapix@hlrs.de>]

Hello,

Thanks for your kind reply.

I am now using StarPU example vector_scal with cpu and gpu implementations. 
But after the execution, I see the data transfer time is always 0. My 
scheduler is "dmda".

What do I need to get the data transfer time measured by StarPU? As I want 
gpu get data first from ram and then put then back to ram, there must be some 
time consumption.

Thanks again.

Best regards,
Fangli 

----- Original Message -----
From: "Samuel Thibault" <samuel.thibault@inria.fr>
To: "Fangli Pi" <hpcfapix@hlrs.de>
Cc: "starpu-devel" <starpu-devel@lists.gforge.inria.fr>
Sent: Wednesday, January 20, 2016 2:58:39 PM
Subject: Re: [Starpu-devel] Questions about StarPU worker recongnization

Fangli Pi, on Wed 20 Jan 2016 14:47:30 +0100, wrote:
> I am interested in how the scheduler choose the worker size?

That depends on the scheduler. Pheft will choose the size which
minimizes the target function (minimum completion time, or combined with
power consumption when that is available). Peager will try to run on the
biggest parallel worker size, without caring about what is currently
available. Of course other schedulers are possible, we have really not
developped those much, just wrote proof-of-concept basic ones.

> But for parallel task, how could I get info about the combined workers? 
> What is the output of the starpu_profiling_task_info : workerid?

Good question, I don't think anybody took care of putting a sensible
information there.

Samuel
-- 
High Performance Computing Center (HLRS)   
University of Stuttgart                    
Nobelstr. 19                               
D-70569 Stuttgart, Germany

phone: ++49-711-685-60442