The GYRE Forums

Hi Rich,
I am totally aware that GYRE uses multi-threading in an efficient way, but I prefer to assign only one CPU to each job, and instead submit a series of jobs using SLURM.
For that, I always set OMP_NUM_THREADS=1
Our IT admin here at IvS noticed a run-time overhead due to the system calling FUTEX a lot of time. Thus, I turned OpenMP off, and recompiled gyre version 2.3, and retried the same jobs. I can tell that I did gain a factor of ~25% speedup.

The reason of this post is only to provide a feedback.

Cheers
Ehsan.

ehsan wrote:Hi Rich,
I am totally aware that GYRE uses multi-threading in an efficient way, but I prefer to assign only one CPU to each job, and instead submit a series of jobs using SLURM.
For that, I always set OMP_NUM_THREADS=1
Our IT admin here at IvS noticed a run-time overhead due to the system calling FUTEX a lot of time. Thus, I turned OpenMP off, and recompiled gyre version 2.3, and retried the same jobs. I can tell that I did gain a factor of ~25% speedup.

The reason of this post is only to provide a feedback.

Cheers
Ehsan.

Hi Ehsan --

Interesting! What system is this running on? I just tried reproducing the problem on Linux (Gentoo, kernel 3.12), and I find no significant difference in runtimes.

cheers,

Rich

Hello Rich,
I logged into one of the compute nodes, and Tell me exactly what specifications you'd like to know and I help you with that.

Cheers
Ehsan.