AW: REMD on IBM iDataPlex DX360M3

From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Thu Jul 03 2014 - 05:38:48 CDT

Sharing GPUs is not a problem. You just need to see every replica as a an individual simulation regarding the number of cores etc. So best choice is to aquire full nodes, as the 1st two processes of each replica (two are required at least btw.) must be on the same physical node. Away from that, every replica behaves like an individual simulation using a part of the resources you gave. So having 16 cores with 8 replica, causes every replica to use 2 cores. So if having multiple replicas per node, the number of replicas per node must be a devider of the number of cores one node contains. If not using any +devices options, the GPUs will allocate for every replica to all GPUs one node contain. Using +ignoresharing to namd2 might bring some speed improvement . Also to prevent random segfaults, as I already posted, change the 1st line in “replica.namd” from

 

replicaBarrier

 

to

 

after 5000 replicaBarrier

 

Im running REMD with GPUs very fine. Once you have understood the principals, its not a problem:

 

1. Make sure 1st two processes of each replica are on the same node (machinefile not unique fashion as its processes round robin until number of aquired cores is reached)

2. Number of replicas per node is a devider of number of cores per node.

3. Number of cores (overall) is devisible by number of replicas and number cores per node

 

Otherwise, give me some error message ;)

 

Norman Geist.

 

Von: Francesco Pietra [mailto:chiendarret_at_gmail.com]
Gesendet: Donnerstag, 3. Juli 2014 11:54
An: Norman Geist
Cc: Namd Mailing List
Betreff: Re: namd-l: REMD on IBM iDataPlex DX360M3

 

I have no recent records. When we tried some months ago, it seemed that the problem is probably trying to make
sure that every replica has the same number of MPI tasks + GPUs. Probably the GPUs can not be shared between replicas. However, I am giving second hand information, probably as bad as any second hand information.

We had no problem on running the same jobs on a BlueGene/Q, except a lot of CPUs required. I hoped to save money through the GPUs (which perform excellently on the same iDataPlex DX360M3 for both normal MD and accelerated MD)

 

francesco pietra

 

On Thu, Jul 3, 2014 at 11:28 AM, Norman Geist <norman.geist_at_uni-greifswald.de> wrote:

What happens if you try?

 

Norman Geist.

 

Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im Auftrag von Francesco Pietra
Gesendet: Donnerstag, 3. Juli 2014 08:15
An: NAMD
Betreff: namd-l: REMD on IBM iDataPlex DX360M3

 

Hello:

Any recent success in carrying out replica exchange molecular dynamics on

IBM iDataPlex DX360M3 by also utilizing the GPUs?
thanks
 
francesco pietra

 

  _____

 <http://www.avast.com/> Fehler! Es wurde kein Dateiname angegeben.

Diese E-Mail ist frei von Viren und Malware, denn der avast! Antivirus <http://www.avast.com/> Schutz ist aktiv.

 

 

---
Diese E-Mail ist frei von Viren und Malware, denn der avast! Antivirus Schutz ist aktiv.
http://www.avast.com

This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:22:34 CST