[petsc-dev] PETSc and threads

[Barry Smith]

> On Jan 8, 2015, at 7:59 PM, Mark Adams <mfadams at lbl.gov> wrote:
> 
> 
> 
> On Thu, Jan 8, 2015 at 7:04 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> > On Jan 8, 2015, at 2:47 PM, Mark Adams <mfadams at lbl.gov> wrote:
> >
> > I have a project that needs to show a plan for moving the the new ORNL machine SUMMIT
> 
>   Isn't Summit IBM PowerN + Nvidia GPUs with a huge percentage of the available floating point performance only available from the GPUs? Thus utilizing the machine means much more utilizing GPUs than utilizing "threads" on the PowerN?
> 
> No, they have very small grids but lots of parallelism because most of the data is in particles.  like 10K particles per cell.  So we just need to run on the CPU.

  Huhh? I don't think one can just run on the CPU's of this machine. You are throwing away 90+ percent of the machine! Did someone in authority tell you it is fine to just use the CPUs in this machine? Hell, I need to talk to them :-)


  Barry

>  
> 
> 
> > and I wanted to see if there is any writeup or data with respect to thread performance in PETSc.
> 
>    We've never run the thread code of PETSc on PowerN (I suspect it would be fairly good but PowerN systems are too expensive to buy to test with :-)).
> 
> Well this is your lucky day!  I should be able to get access to the P9 of SUMMIT :)