On 2007-04-04, Michael E. Thomadakis <miket@sc.tamu.edu> wrote:
>
> More specifically, I would expect that SMT ON would benefit workload
> mixes with high thread concurrency but with low contention on common
> thread resources (L1/L2/L3 and functional units). For instance, I would
> expect a cluster with a large number of distributed (and multithreaded)
> processes running to benefit from SMT ON: twice as many h/w threads can
> be active at a time, obviating the need for costly thread-switch.

I did some benchmarking on a P5+ cluster last fall, and was
quite surprised that SMT=ON (and 1 task per virtuall cpu vs.
1 task per physical cpu in SMT=OFF) was giving best performance
on all the benchmarks we did. Even for 16 node (~250 task)
mpi-jobs.


-jf