Sites Inria

Version française

Inria-Urbana Champaign

Amélie Castan - 30/09/2011

Thom Dunning, a successful collaboration

Thom Dunning

On June the 27th, The Joint Laboratory for Petascale Computing (JLPC), formed by the University of Illinois and Inria, held its fifth workshop in Grenoble. Thom Dunning, director of the Institute for Advanced Computing Applications and Technologies and the National Center for Supercomputing Applications (NCSA) at the University of Illinois, explains why this collaboration is a success.

You started JLPC lab with Inria in 2009. How does this collaboration take place?

Thom Dunning:  A very important aspect of the collaboration is the joint workshops we hold every 6 months. They are primarily focused on reporting on progress that has been made in existing projects and looking for new opportunities. And new opportunities arise all the time as we better understand the issues associated with petascale computing. The other major activity is the exchange of students and senior researchers in Illinois and in the various Inria laboratories in France. They can range from 3 days to weeks and months.  Joint teams call each other quite frequently and work on joint publications, joint presentations at various conferences, develop software made available for the larger scientific community... And we often involve other research groups, for example in Japan or in France.

This year, the JLPC launched the G8 “Enabling Climate simulation at extreme scale” project. How did it start?

Thom Dunning:  At the end of 2010 we started discussing the issues associated with climate modeling. Climate modeling has always taxed the most powerful computers we have because the models are very complex and involve coupling many different components together, with unique computer science issues associated with those simulation tasks. This discussion that came up is a typical way that collaborative projects start with: somebody has an idea, it is discussed with colleagues, the discussion gets deeper and deeper and finally a research pathway opens up.

Since the beginning of the collaboration with Inria through the JLPC Lab, which results did you get?

Thom Dunning: This joint work is addressing a richer set of problems than I had originally anticipated: fault tolerance and resilience, optimizing numerical libraries for petascale and exascale computing, new programming models and run-time environments for Blue Waters, the petascale computing system that will be installed beginning 2012 at the National Center for Supercomputing Applications, in Illinois.

The problems we tackle are so difficult that we need people with a broad range of expertise and different viewpoints.

What new opportunities have arisen for the last workshop in June?

Thom Dunning: One of the problems that have arisen these last years is: how can we improve the efficiency of I/O (input/output to disks)? Blue Waters will have about 20 petabytes of disk storage. Moving 20 petabytes of data back and forth can be a very time consuming task. An Inria/Illinois joint research project was recently established in this area.

With Blue Waters, will you continue the collaboration with Inria?

Thom Dunning: Yes. Some of the ideas of the joint lab are very good, but testing at scale really needs to be done. We are also looking at the technologies that will be used for exascale computing, for example GPUs: simplified types of processors that can execute instructions much faster than traditional CPUs. There is quite a bit of work to do on an application to enable it to make full use of the architecture of the GPU.

What is the contribution of Inria teams to your research?

Thom Dunning: Their work is both complementary and supplementary. The problems we tackle are so difficult that we need people with different expertise, experience and viewpoints. The bridges we build between the experience Inria teams have, especially in numerical libraries and resilience, and our expertise in programming models, communication libraries, I/O and GPU create a very rich knowledge environment.

Keywords: Modélisation climatique Blue Waters GPU