[HPGMG Forum] Execution heterogeneity

Sam Williams swwilliams at lbl.gov
Wed Apr 30 18:28:57 UTC 2014


Let me assume a host+accelerator model.  My approach would be coarse grained.  Each level can run either on the host or an accelerator.  This can be determined statically tuned at run time.  Simultaneously heotergeneous implementation (i.e. within a level some fraction of the cells are on the host and some fraction on the device) seems like a nightmare to program.



On Apr 30, 2014, at 11:24 AM, Brian Van Straalen <bvstraalen at lbl.gov> wrote:

> That is a whole other growing issue, algorithm heterogeneity.  I’m thinking architecture.
> 
> Latency cores vs throughput cores, partitioned local memory, co-processing.
> 
> 
> On Apr 30, 2014, at 12:52 PM, Jed Brown <jed at jedbrown.org> wrote:
> 
>> Brian Van Straalen <bvstraalen at lbl.gov> writes:
>> 
>>> I’m wondering where users can insert heterogeneity in the
>>> execution…which seems a big coming issue.
>> 
>> You mean unavoidable/unpredictable divergence between vector lanes (such
>> as appears with iterative solves in Riemann solvers, material models,
>> etc)?  I wrote my thoughts on that here:
>> 
>>  https://hpgmg.org/fe/#linear-or-nonlinear
>> 
>> I'd love to hear ideas about how to incorporate this into a benchmark.
> 
> Brian Van Straalen         Lawrence Berkeley Lab
> BVStraalen at lbl.gov         Computational Research
> (510) 486-4976             Division (crd.lbl.gov)
> 
> 
> 
> 



More information about the HPGMG-Forum mailing list