Re: Seward I/O

Posted by dennis mueller on October 17, 2013 at 17:02:12:

In Reply to: Re: Seward I/O posted by Roland Lindh on October 17, 2013 at 16:42:17:

: Hi Dennis,
: the OrdInt(X) files in each of the temp_(x) folders are different - if you run in parallel. Each core is computing a different set of shell-quadruplets, hence they all will have different OrdInt(X) files. If all the cores share the same I/O bus, you are in problem. Could this potentially be the problem - the I/O does not scale with the number of cores?

: Roland

Hi,

thank you for your fast response!
i am afraid this is exactly the problem. i am running molcas in parallel mode on one local machine with 16 physical cores. there is no distribution to other machines over LAN as of yet. so if i use multiple cores on one machine they inevitably use the same I/O bus which then generates this bottle neck and makes the seward module so slow. i am afraid there is no workaround for that, other than using molcas on a cluster with not too many cores per node, or is there?
is this the setup you would recommend for molcas instead of one big machine with a high number of cores?

i am very grateful for your help.
Kind Regards, Dennis.

Follow Ups:

Re: Seward I/O Steven 08:05:28 6/23/14 (1)
- Re: Seward I/O Ruihong Duan 11:15:03 9/02/14 (0)

Post a Followup

Name:
E-Mail:

Subject:

if B is 1s22s22p1, what is Li?

Passfield:

Comments:
: : Hi Dennis, : : the OrdInt(X) files in each of the temp_(x) folders are different - if you run in parallel. Each core is computing a different set of shell-quadruplets, hence they all will have different OrdInt(X) files. If all the cores share the same I/O bus, you are in problem. Could this potentially be the problem - the I/O does not scale with the number of cores? : : Roland : Hi, : thank you for your fast response! : i am afraid this is exactly the problem. i am running molcas in parallel mode on one local machine with 16 physical cores. there is no distribution to other machines over LAN as of yet. so if i use multiple cores on one machine they inevitably use the same I/O bus which then generates this bottle neck and makes the seward module so slow. i am afraid there is no workaround for that, other than using molcas on a cluster with not too many cores per node, or is there? : is this the setup you would recommend for molcas instead of one big machine with a high number of cores? : i am very grateful for your help. : Kind Regards, Dennis.

[ Follow Ups ] [ Post Followup ] [ Molcas user's WWWBoard ]