Posted by dennis mueller on October 16, 2013 at 14:23:04:
I have molcas running on a local cluster with 16 physical cores (32 with HT) and 96Gb RAM and 500Gb SSD scratch space (the scratch space is small but sufficient. i don't run into out of disk errors). I use openmpi for parallelisation and the acml math lib.
i am encountering a problem with the seward module i was not expecting on a local system. when calculating the 2electron integrals seward is unbearably slow due
to I/O. I know this is a common problem and i did some reading in the billboard and the manual already. i think the main problem is that the *.OrdInt(x) files are not only in the top level workdir but also in
every tmp_(x) subfolder redundantly. these are not links! not even hardlinks. since there are a lot of these files depending on the basis this naturally boosts the I/O through the roof. sorry for this long
introduction to my problem i hope you are still sticking with me. how can i tell molcas (or mpi) to use links to these files instead of physically copying them? this would not only reduce diskspace but also
speed up the calculations.
i have a second question which concerns the molcasmem variable. i don't know if this is intended or not but if i set CPUS=4 for example and MOLCASMEM=8000 then 32 GB of RAM are allocated. so molcasmem defines the max amount of mem used by one instance of the seward module? is this correct and intended?
thank you for your help i very much appreciate it. if you need any further information on my system or a testinput or anything else, please let me know.
Regards, Dennis Müller.
Post a Followup