[HPGMG Forum] Do we want the benchmark to go into intrinsics?

Sam Williams swwilliams at lbl.gov
Tue Apr 29 17:14:48 UTC 2014


optimized = simd intrinsics, array padding to avoid cache aliasing cognizant of line size, associativity, cache size, replacing omp workshare with static parallel regions and manually fusing operations/communication together, etc...

I think you can have a couple reference implementations...
- C with MPI and/or OpenMP (so you can run on distributed memory machines or SGI UV ccNUMA)
- CUDA with MPI (little hesitant on this given there is either a new version or new release candidate of CUDA every 6 months)

The kiviats were constructed with code as is.  HPCG defenders promise better performance with optimization.
Showing HPGMG optimized performance and performance characteristics looks better than vanilla code/HPCG doesn't address the argument.  It only says optimization is beneficial.

I think the reference implementation should be "good".  I'm much more curious as to why xlc fails so badly on HPGMG-FE while it does pretty good on HPGMG-FV.



On Apr 29, 2014, at 10:06 AM, Mark Adams <mfadams at lbl.gov> wrote:

> 
> 
> 
> On Tue, Apr 29, 2014 at 12:53 PM, Sam Williams <swwilliams at lbl.gov> wrote:
> Were any of the codes Bert/Ian analyzed manually simdized for QPX ?
> 
> I want to keep reference separate from optimized.  
> 
> What do you mean by optimized?
> 
> Are you saying that you want the reference implementations to be straight C/OpenMP/MPI/CUDA?
> 
> That is a defensible position.  I am just asking.
>  
> I didn't want to manually simdize anything until the community accepted HPGMG as a useful benchmark and agreed on the definition.  I also plan on keeping any of my optimizations private until published.
> 
> 
> Can you explain your reasons for this?  Do you think the community would be less likely to accept HPGMG if they saw a lot of optimized code in the repo?
> 
> Do you agree with my statement in my last email that the reference implementation should be pretty good for a least one platform?  I thought we were on the same page on this but if not we should do that now.
> 



More information about the HPGMG-Forum mailing list