Re: AW: Simultaneous calculation on CPU-only nodes and CPU/GPU node (with or without rCUDA)

From: Benjamin Merget (benjamin.merget_at_uni-wuerzburg.de)
Date: Sun May 27 2012 - 09:04:48 CDT

Hi Norman,

Thanks for your reply! Unfortunately, I couldn't get it running the way
you suggested. I tried it with the ibverbs-smp-CUDA binaries and got the
following errors:

Charmrun> error 93620 attaching to node:
Socket closed before recv.

My own MPI-version was not working either, since its charmrun works
differently.

I tried another approach, which was installing the CUDA and non-CUDA MPI
versions locally to the nodes in the same paths. So that the different
versions are executed by mpiexec on the different nodes (which is
basically what we tried before, but now via MPI, I suppose...).
Unfortunately, this wasn't working either. Here, the program crashed
with the error

Unknown compute type in ComputeMgr::createCompute()

for each process which was supposed to run on the CPU node.

I didn't think that my problem was something so exotic...
I always imagined this to be a pretty standard task.

Do you have any other suggestions?

Thanks very much,
Benjamin

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:21:34 CST