AW: vmd-l: Building NAMD 2.12 from Source with IB and CUDA

From: Greipel.Joachim_at_mh-hannover.de
Date: Thu Mar 02 2017 - 05:53:58 CST

Hi,

 

yes, this question should indeed have been placed at the namd mailing list.
Mea culpa.

 

I just found the solution to the compiling problem myself by invoking
“./build” without options. After answering some simple questions I ended up
with the commandline

 

“./build charm++ verbs-linux-x86_64 smp” –with-production”

 

With this charm architecture I managed to run “configure” and the rest of
the build process. But when invoking the program I found that the binary of
namd 2.12 is still dependent on compute capabilities > 3, so I stopped here.

 

Best wishes,

 

Joachim Greipel

 

 

 

Von: Norman Geist [mailto:norman.geist_at_uni-greifswald.de]
Gesendet: Donnerstag, 2. März 2017 11:38
An: Greipel, Joachim Dr.
Cc: vmd-l_at_ks.uiuc.edu
Betreff: AW: vmd-l: Building NAMD 2.12 from Source with IB and CUDA

 

Hi,

 

NAMD’s distributed memory parallelization is based on the charm++
programming model (and library). This library has its own network
communication layer but can also use MPI as its backend. In the past this
library needed to be compiled with MPI support for some of NAMD’s multi-copy
features (e.g. REMD). Now this has changed and the NAMD people say that you
should not use an MPI based charm++ build when using CUDA too, anymore as
the performance shall be better with pure charm++.

 

So either build charm++ without MPI support, or simply remove the check from
the build script (that’s what I did), but this might stop working with
upcoming namd version.

 

Best wishes

 

Norman Geist

 

Btw. this question should have been placed at the namd mailing list
probably.

 

 

Von: owner-vmd-l_at_ks.uiuc.edu [mailto:owner-vmd-l_at_ks.uiuc.edu] Im Auftrag von
Greipel.Joachim_at_mh-hannover.de
Gesendet: Donnerstag, 2. März 2017 09:57
An: vmd-l_at_ks.uiuc.edu
Betreff: vmd-l: Building NAMD 2.12 from Source with IB and CUDA

 

Hi,

 

I got two machines connected by Infiniband, running CentOS 7.3, using 8
physical cores each and harboring two QuadroFX 3800 each.

The last NAMD Version I could successfully run NAMD using both CUDA and
Verbs was 2.10 (Linux-x86_64-ibverbs-smp-CUDA Binary) because the compute
capabilities of the cards is only 1.3.

With the latest Version (2.12) the requirements are compute capabilities >3,
so this version does not work.

So I tried to build NAMD 2.12 from Source for ibverbs and smp, using cuda
toolkit 6.5, but failed:

I successfully built NAMD 2.12 from Source with ibverbs OR with CUDA. But if
I tried to build it with ibverbs AND CUDA support, I got the error messages
from configure:

 

----------------------------------------------------------------------------
-------------------------------------------------

./config Linux-x86_64-g++ --with-cuda --cuda-prefix /usr/local/cuda
--charm-arch verbs-linux-x86_64

 

ERROR: CUDA builds require non-MPI SMP or multicore Charm++ arch for
reasonable performance.

 

Consider ibverbs-smp or verbs-smp (InfiniBand), gni-smp (Cray), or multicore
(single node).

----------------------------------------------------------------------------
-------------------------------------------------

 

When I go into the charms subdirectory and invoke ./build –help to scan the
possible charm architectures, I found nothing that fulfils these
requirements.

 

What can I do to get the 2.12 Version running on the configuration above?

 

Joachim

 

 

--
Dr. rer. nat. Joachim Greipel 
 
Med. Hochschule Hannover 
Biophysikalische Chemie OE 4350 
Carl-Neuberg-Str. 1 
30625 Hannover 
Germany 
Fon: +49-511-532-3718 
Fax: +49-511-532-16-3718
E-Mail:  <mailto:greipel.joachim_at_mh-hannover.de>
greipel.joachim_at_mh-hannover.de
 

This archive was generated by hypermail 2.1.6 : Sun Dec 31 2017 - 23:21:07 CST