This is not surprising. I don't see what GPUs you used, but with the
latest Volta lines it's not possible to get better performance than the
single GPU case unless you have some really specialized interconnects and
like eight of them--a corner case that's not worth replicating. Just stick
to single GPU runs. The multi-GPU stuff remains in the code because we
will not break functionality that was added years ago, and we may return to
the multi-GPU world in time, but for now the benefit of running multiple
GPUs on the Pascal architectures is marginal and Volta is nonexistent.
Dave
On Fri, Jun 22, 2018 at 7:59 AM 陈金峰 <201612095.mail.sdu.edu.cn> wrote:
>
> Hello amber users:
> I am starting to use a station with 3GPUs, I used the following
> command :
> nohup mpirun -np 3 pmemd.cuda.MPI -O -i nvt.in -o nvt.out -r...
> It did work ,but it performed even slower than ran on only one GPU, and
> the
> output information is as follows:
>
> [[9951,1],2]: A high-performance Open MPI point-to-point messaging
> module
> was unable to find any relevant network interfaces:
> Module: OpenFabrics (openib)
> Host: hj191
> Another transport will be used instead, although this may result in
> lower performance.
> NOTE: You can disable this warning by setting the MCA parameter
> btl_base_warn_component_unused to 0.
>
> --------------------------------------------------------------------------
> [hj191:04235] 2 more processes have sent help message
> help-mpi-btl-base.txt
> / btl:no-nics
> [hj191:04235] Set MCA parameter "orte_base_help_aggregate" to 0 to see
> all
> help / error messages
> Note: The following floating-point exceptions are signalling:
> IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
> Note: The following floating-point exceptions are signalling:
> IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
> Note: The following floating-point exceptions are signalling:
> IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
> So what's the problem? what should I do to get a proper performance?
> Any suggestions will be appreciated!
> Thank you!
> _______________________________________________
> 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 Fri Jun 22 2018 - 07:30:02 PDT
