[Beowulf] ***UNCHECKED*** HPCG

[Richard Walsh]

Prentice wrote:

> When I compare my HPL results to my HPCG results, I'm getting HPCG
> results that are 0.3 - 0.5% of HPL. On the HPCG Top500 list, most
> systems are getting 2-3% of HPL, so I'm off by an order of magnitude.

Of course HPCG is a bandwidth limited application so it will never come
close to HPL.  It is likely that you are seeing the effect of running the
stock
3.1 release and comparing that to optimized releases from Intel and
elsewhere
that are much faster, and faster still if they are most running on the GPU.

You can get better CPU performance using Intel optimized version distributed
with the cirompiler, or grabbing NVIDIA's version for the GPU.   I know
that a
fully optimized version of the code runs about 50% faster than the
stock 3.1
version no matter how you compile it.  You should also look at not running
fully packed and play with the rank to thread ratio.  The idea is to get as
close
to theoretical bandwidth from the node as possible and this may mean running
unpacked.

Cheers!

rbw

On Wed, Aug 5, 2020 at 1:09 PM Prentice Bisbal via Beowulf <
beowulf at beowulf.org> wrote:

> Beowulfers,
>
> Do any of you have any experience using HPCG as a benchmark. I'm trying
> to compare the performance of several different processors for an
> upcoming purchase. I've already run LINPACK, and now I'd like to run
> HPCG. It seems the only tuning parameter is the size of the local grid
> in the x,y,z dimensions.
>
> While the guidelines say to increase the gridsize until the job consumes
> 1/4 or more of RAM, my testing has shown that as the gridsize goes up,
> so does the performance,  and it keeps going up for me until I consume
> all the memory and the job gets killed by Slurm for exceeding memory
> requirements.
>
> I've been doing a lot of Google searching for how to tune HPCG for
> maximum results, and there are some papers for tuning HPCG for large
> supercomputers. In these cases, they use x,y,z dimensions that are not
> necessarily equal, but I don't understand how they determined to use
> these unique values for x,y,z.
>
> When I compare my HPL results to my HPCG results, I'm getting HPCG
> results that are 0.3 - 0.5% of HPL. On the HPCG Top500 list, most
> systems are getting 2-3% of HPL, so I'm off by an order of magnitude.
>
> --
> Prentice
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
> https://beowulf.org/cgi-bin/mailman/listinfo/beowulf
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://beowulf.org/pipermail/beowulf/attachments/20200806/73ab4033/attachment.html>