jcduell_at_lbl_dot_gov
Date: Tue Dec 06 2005 - 10:06:33 PST
On Tue, Dec 06, 2005 at 10:26:55AM -0500, Eric Frederich wrote:
> I plan on doing some benchmarking of various parts of UPC. My test setup at
> work consists of about 40 or so Linux workstations each with two Opteron
> processors.
> I am wondering what is the best way to run parallel programs on these
> machines?
>
> Should I list each workstation twice in $UPC_NODES or just once?
> Should I pass -pthreads to upcc to enable pthread support?
> What aguments should I use with upcrun (-n -nodes -c and -p)?
Generally, the fastest way to run on a cluster of SMPs is to run 1
process per machine, using -pthreads. That way you get shared memory
between UPC threads that live on the same node.
So, for a 3-node cluster of 2-way SMPs, try
upcc -pthreads=2 foo.upc
export UPC_NODES="node1 node2 node3"
upcrun -n 6 a.out
Note that the default value of -pthreads is 2, so it's superfluous here,
but note the general idea (that you can embed a default number of pthreads
per process into the UPC executable). You could also pass '-p 2' to
upcrun to set the number of pthreads per process at run time.
There are some issues with using -pthreads that you may run into. Most
commonly, some math/science libraries aren't pthread-safe, and will
break under -pthreads. See the section "Using pthreaded Berkeley UPC
programs" in our User's Guide. If you run into these issues, just use 1
UPC thread per process:
upcc foo.upc
export UPC_NODES="node1 node1 node2 node2 node3 node3"
upcrun -n 6 a.out
Cheers,
--
Jason Duell Future Technologies Group
<jcduell_at_lbl_dot_gov> Computational Research Division
Tel: +1-510-495-2354 Lawrence Berkeley National Laboratory