Hi Eduardo,
A few things to note here. Firstly you are unlikely to see good scaling between your nodes unless they have very high end connectivity (e.g. infiniband) between them and even then only for CPU runs. As such I would not try to run that benchmark script across nodes. It was never designed for that. Thus you should simply set CPU_COUNT=12 and GPU_COUNT=2 - that will run each benchmark in the following ways:
1) Using 12 CPU cores on a node
2) Using GPU 0 on a node
3) Using GPU 1 on a node
4) Using GPU 0 and GPU 1 to each run a separate benchmark - performance should match items 2 and 3 above. If it doesn't then your cooling is insufficient.
5) Using GPU 0 and GPU 1 to jointly run the benchmark - on K20s you may still see some improvement here but on more modern faster GPUs you are unlikely to see performance improvement by using more than 1 GPU at once for a single calculation (the exception being large > 5K atom GB calculations).
Note when running on GPUs the CPU count is irrelevant - 1 CPU core is used per GPU and it's job is primarily just I/O.
I hope that helps
All the best
Ross
> On Jan 29, 2019, at 12:34, Eduardo Almeida Costa <eacosta.uesc.br> wrote:
>
> Hello everyone.
>
> I have available 3 nodes with 12 CPUs and with 2 Tesla K20 GPUs each.
> (Total 36 CPUS and 6 GPUs)
>
> I'm using AMBER 14 + AmberTools 15.
>
> I downloaded the AMBER 14 benchmark files (
> http://ambermd.org/gpus14/benchmarks.htm).
>
> My question is about the correct configuration of the parameters
> "CPU_COUNT" and "GPU_COUNT" in the file "run_bench_CPU + GPU.sh".
>
> Should I put how much each node has, or put the total?
>
> # ------ SET FOR YOUR SYSTEM -------
> #
> GPU_COUNT = 2
> CPU_COUNT = 12
> #Should we run large STMV 1 million atom + test?
> RUN_LARGE_BENCH = true
> #
> # ----------------------------------
> Or
> # ------ SET FOR YOUR SYSTEM -------
> #
> GPU_COUNT = 6
> CPU_COUNT = 36
> #Should we run large STMV 1 million atom + test?
> RUN_LARGE_BENCH = true
> #
> # ----------------------------------
>
> My best regards.
>
> Eduardo.
> --
> Eduardo Almeida Costa
> Analista Universitário - 73.528.957-1
> NBCGIB/UESC
> Tel: +55 73 3680 5212
> Bel. em Ciência da Computação/UESC
> Me. em Genética e Biologia Molecular/UESC - CAPES 5
> Dr. em Biologia e Biotecnologia de Microrganismos/UESC - CAPES 4
> Programa de Doutorado Sanduíche no Exterior 2017 (PDSE 2017) - Aberystwyth
> University, País de Gales, Reino Unido, com apoio da Coordenação de
> Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Código de
> Financiamento 001.
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jan 30 2019 - 08:00:02 PST
