[Beowulf] Best case performance of HPL on EPYC 7742 processor ...

[Scott Atchley]

I do not have any specific HPL hints.

I would suggest setting the BIOS to NUMAs-Per-Socket to 4 (NSP-4). I would
try running 16 processes, one per CCX - two per CCD, with an OpenMP depth
of 4.

Dell's HPC blog has a few articles on tuning Rome:

https://www.dell.com/support/article/en-us/sln319015/amd-rome-is-it-for-real-architecture-and-initial-hpc-performance

Scott


On Fri, Aug 14, 2020 at 5:30 PM Richard Walsh <rbwcnslt at gmail.com> wrote:

>
> All,
>
> What have people achieved on this SKU on a single-node using the stock
> HPL 2.3 source... ??
>
> I have seen a variety of performance claims even as high as 90% of its
> nominal
> per node peak of 4.608 TFLOPs.  I can now get above 80% of peak, but not
> higher.
> I have heard that to get higher values special BIOS settings are required,
> including
> the turning off SMT which allows the chip to turbo higher.  Remember this
> is not the
> 7542 processor with 32 cores per chip and the same bandwidth per socket as
> the
> 7742 which can turbo to over 100% of nominal peak for HPL.
>
> If people have gotten higher single node numbers ... what is your recipe
> ... ??
>
> I am particularly interested in BIOS settings, and maybe surprise settings
> in the HPL.dat file.  Do higher performing runs require using close to the
> maximum memory on the node ... ??  As this is single-node, I would not
> expect choice of MPI to make a difference
>
> To get to 80% with SMT on in the BIOS, I am building with an older Intel
> compiler and MKL that still recognizes the MKL_DEBUG_CPU_TYPE=5.
> Running so that the number of MPI ranks run on the node matches the
> number of CCXs seems ot give the best numbers.
>
> Following the tuning instructions from AMD for using BLIS and GCC for
> the build does not get me there.
>
> Thanks,
>
> Richard Walsh
>
> _______________________________________________
> 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/20200817/85d3d4a0/attachment.html>