Ppl-accel mailing list

[Ronak Buch <rabuch2 AT illinois.edu>]

---------- Forwarded message ----------
From: Kale, Laxmikant V <kale AT illinois.edu>
Date: Fri, Sep 1, 2017 at 3:27 PM
Subject: FW: Urgent --- NAMD performance on GPU machines
To: "Buch, Ronak Akshay" <rabuch2 AT illinois.edu>

 

From: Emad Tajkhorshid <emad AT life.illinois.edu>
Date: Friday, September 1, 2017 at 3:14 PM
To: "Hardy, David Joseph" <dhardy AT illinois.edu>, Jim Phillips <jim AT ks.uiuc.edu>, "Kale, Laxmikant V" <kale AT illinois.edu>
Cc: Emad Tajkhorshid <emad AT life.illinois.edu>, John Stone <johns AT ks.uiuc.edu>, Aleksei Aksimentiev <aksiment AT illinois.edu>, "Luthey-Schulten, Zaida Ann" <zan AT illinois.edu>
Subject: Urgent --- NAMD performance on GPU machines

 

Dear all, 

 

I would like to bring this important issue to the attention of all major parties involved urging everyone to allocate time and resources to address it as urgently as possible. We have been paying attention, increasingly, to the fact that with the availability of faster GPUs and nodes with multiple GPUs, there are some major bottlenecks in NAMD that need immediate attention. As discussed in the last meeting, GPU performance has become a major flag for our performance arch-competitor, AMBER, and a significant number of users migrating to it just because of that. More recently, there have been more pressing reasons to take care of the new bottlenecks when running NAMD on GPU enabled platforms. A major one is the need for new benchmarks for NVIDIA for the new Volte GPU and Summit. We have all reached the conclusion that we have to immediately take care of the major bottlenecks to be able to provide better benchmarks for NAMD runs on these novel architectures. 

 

NVIDIA is highly engaged and heavily vested in this aspect. There are two programmers on their end who continue to work on the problem. On our end, Dave has been spearheading the effort and started to work on profiling and identifying the bottleneck more rigorously. It seems that at least one of the main steps to be taken in to implement the integrator on the GPU, which is something Dave is looking into as the first step. 

 

I would like to urge all to actively engage and participate in this important step. In particular, active participation of Sanjay’s people in the effort appears to be essential. Sanjay, would you kindly include people in your group who are relevant to this project in the email. Dave, you should let Sanjay’s student know and ask for their help if there is something they can do. 

 

Thank you all. I reiterate that this aspect of NAMD is critical for our competitiveness in the market and we have to focus on it in major ways. 

 

Thanks, and wish you all a great long weekend. 

- Emad

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

Emad Tajkhorshid, PhD
University Scholar
J. W. Hastings Professor of Biochemistry, Biophysics, and Computational Biology
Group Leader, Computational Structural Biology and Molecular Biophysics Group
Director, NIH Center for Macromolecular Modeling and Bioinformatics 
Director, NSF nanoBIO Node

Department of Biochemistry, School of Molecular and Cellular Biology
Beckman Institute for Advanced Science and Technology and 
Center for Biophysics and Computational Biology
University of Illinois at Urbana Champaign

Office: 3009 Beckman Institute
Tel: 217-244-6914
Email: emad AT life.illinois.edu
web: http://csbmb.beckman.illinois.edu

Mailing address: 
Beckman Institute, 
405 N. Mathews Ave., 
Urbana, IL 61801, USA