Thank you, Dr. Plimpton.
I am using our University cluster, which runs Linux. The nodes I am using typically has up to 64 cores/node. I am not locking my processes to any cores (actually I just looked it up to know how it is done). Like you said, the problem is probably with the way the job is scheduled on my machine, but what I don’t understand is that, the problem only occurs when I use fix shake.
Cheers,
Vikram Reddy Ardham
but what I don’t understand is that, the problem only occurs when I use fix shake.
Sorry, I have no idea.
Steve
Thank you, Dr. Plimpton.
I am using our University cluster, which runs Linux. The nodes I am using typically has up to 64 cores/node. I am not locking my processes to any cores (actually I just looked it up to know how it is done). Like you said, the problem is probably with the way the job is scheduled on my machine, but what I don’t understand is that, the problem only occurs when I use fix shake.
You need to provide convincing evidence to support such a statement and that can be only obtained in cooperation with your local system managers.
Have you made some scaling tests proving, that you are using a suitable number of nodes/processor cores?
Can you verify that all “bad” jobs have run on different subsets of nodes and that the “good” and “bad” jobs were all run on comparable hardware with no other jobs sharing your nodes?
Axel