| 12:32:25 PM PDT - Mon, Oct 6th 2014 |
|
Zontang,
My advice is that, for systems similar to cascade, you should be using eight (or even four) processes on each node. This allows you both to have more memory available for each process and you would get better communication performance because fewer processes are sharing the same network interface.
"ccsd_sxy: write failed" failure: you are probably filling up the /scratch filesystem on cascade. My suggestion is to use the direct approach ("nodisk" keywork in the CCSD input section)
"g_st2" warnings: the warning message is advising you to add another 3.3GB of memory of MA a.k.a. stack memory.
I strongly encourage you to do that since the parallel performance improves dramatically. You might have to use just 4 cores/node to accommodate this memory request.
|