[slurm-users] unable to run on all the logical cores

[Diego Zuccato]

Il 12/10/20 05:55, William Brown ha scritto:

> We always disable hyper-threading on the compute nodes, at the
> recommendation of the company that installed it, who reported that it
> ran faster that way.  Multi-threading has a cost and is not ideal for
> compute workloads.  It is better for things like web servers where some
> tasks are waiting IO or user input.  That said, it likely depends on the
> CPU and what is true for Intel might not be true for AMD.
Well, we did some tests too. Mostly w/ MPI on the newer nodes (2x14x2).
We didn't test disabling HT, just "not using" (scheduling a single
process per core). Even FPU-intensive jobs (where there should be
maximum contention of the FPU between the two threads) seems to be
worthwile to use HT.
Say a job can process 100 iterations when using a single thread on a
reserved core. If I let it use both threads it can process roughly 180
iterations. So every thread is actually about 10% slower, but overall I
can get 1.8 times the throughput when using HT.

That's just my very limited experience, but it seems it could be
worthwhile to test.

-- 
Diego Zuccato
DIFA - Dip. di Fisica e Astronomia
Servizi Informatici
Alma Mater Studiorum - Università di Bologna
V.le Berti-Pichat 6/2 - 40127 Bologna - Italy
tel.: +39 051 20 95786