CUDA Spotlight: Erik Lindahl
GPUs for Molecular Dynamics
This week's spotlight is on Dr. Erik Lindahl of KTH Royal Institute of Technology and Stockholm University.
Erik is a project leader for GROMACS, a popular open-source molecular dynamics program.
This interview is part of the CUDA Spotlight Series.
Q & A with Erik Lindahl
NVIDIA: Erik, in a nutshell, what is GROMACS?
Erik: GROMACS (GROningen MAchine for Chemical Simulations) is a molecular dynamics package primarily designed for simulations of proteins, lipids and nucleic acids. We originally developed it at the University of Groningen. It is now maintained by contributors in universities and research centers across the world. It is free, open source software released under the GNU General Public License.
NVIDIA: You have a new release coming up, correct?
Erik: Yes. We are close to releasing Gromacs 4.6, which is the first version where CUDA is part of our main acceleration strategy and parallelization rather than "bolted on" as a separate module. We are solid CUDA evangelists and see GPUs as natural co-processors for use on some of the largest supercomputers in the world.
NVIDIA: Who should consider downloading GROMACS?
Erik: For obvious reasons, we believe it's a very useful toolkit for biomolecular simulation - that is, it is a very fast (and free!) program to predict how molecules such as the ion channels cause your heartbeats to occur on microsecond scale.
For a more general audience, I think it is a good concrete example of an exceptionally portable and highly accelerated parallel scientific computing application.
In fact, one of our toughest challenges in getting impressive acceleration on NVIDIA cards has been the decade of low-level SSE, AVX and multi-threaded tuning that has gone into the code, and it is only in the last two years that the GPUs are now clearly outperforming even expensive multi-socket CPU nodes for simulations. [Editor's note: To try out GROMACS on a remotely-hosted GPU cluster, see www.nvidia.com/GPUTestDrive]
NVIDIA: Tell us about the GROMACS team.
Erik: GROMACS is a team effort organized from Stockholm, but I can't even take credit for our current GPU code - that amazing work is entirely due to Szilárd Páll (student) and Berk Hess (associate professor) in the lab.
Berk is a walking encyclopedia of mathematics and simulation algorithms (not to mention implementations,) but none of us would have been able to accomplish this without Szilárd - it never ceases to amaze me how he refuses to accept "good enough" but always manages to squeeze about another 10, 20, 30, and even 40% performance long after Berk and I think we've reached the limit...
I think the unifying motivation for most of us is that we originally started implementing features, parallelization, and acceleration that we needed to accomplish our application work. However, somewhere on the road you suddenly get stuck and realize how fascinating it is to enable computers to do things they have never been able to do before, and then parallel scientific software becomes an amazing adventure.
NVIDIA: What are you currently working on, in terms of research?
Erik: On the biological side, we're working a lot with pentameric ligand-gated ion channels. That might not sound sexy, but it is. Every time you have a drink (or cigarette), small molecules such as ethanol (or nicotine) bind to these receptors and affect nerve signals, which is largely what causes that pleasant feeling. For the first time ever, we can now use computer simulations to predict exactly how these molecules work!
NVIDIA: When did you first learn about GPUs?
Erik: Ian Buck (now at NVIDIA) was doing his Ph.D. at Stanford while I was a postdoctoral scholar there, and he had this really crazy idea about doing a molecular simulation on a GeForce4 graphics card in collaboration with us and Vijay Pande of Folding@Home. You might not remember, but in those days there wasn't even full floating-point support, and CUDA wasn't around yet either. Despite lots of issues the concept actually did work, and despite challenges along the way, we have not looked back since.
NVIDIA: What kind of advantages have you achieved with CUDA?
Erik: In light of our early trial-and-errors, the main difference is that CUDA is a long-term stable development platform that we can trust. That would obviously not be worth anything if it didn't also provide amazing performance, but considering the amount of work we need to invest in the code, it is critical for us that we can reuse it on future platforms rather than start over from scratch.
In practical terms, CUDA has made it possible for us to go from tweaking code to gain a few percent of improvement on CPUs (which is a lot of money when you use a supercomputer for months!) to getting several hundred percent speed-up.
NVIDIA: How did you first get interested in molecular dynamics?
Erik: I've been fascinated by computers, physics and biology since I was a little kid, and one of the huge agonies when I wanted to do a PhD was the realization I would have to choose between them. Luckily enough, I stumbled upon an advisor who was working on esoteric methods to study biological molecules on computers using physical principles. I had no idea how small the field was at the time, but I think it shows how difficult it can be to predict the future - 15 years later it is a very hot field with huge interest from experimental biochemists.
NVIDIA: What do you like to do when you are not working?
Erik: Actually, I still don't feel like I have a job, but that I'm getting paid to think and play all day - academia is a wonderful place to really explore ideas! However, I'm also a big opera fan, and Stockholm is a great place for sailing.
NVIDIA: What is the technological achievement you are most proud of?
Erik: I'm not sure I could mention a single one - all the technological advances are really fun short-term when you are nailing that difficult algorithm, but long-term I think we're most proud of our small contribution to establish a new field of research where thousands of non-programmer colleagues use our code to accomplish simulation results that they could not even do in an experimental laboratory a decade ago!
NVIDIA: What's next in the field of bioinformatics?
Erik: We are already seeing a rapid trend with computer simulations replacing “wet lab” experiments - even in the field of bioinformatics - and I think this will accelerate even further. Just as the airplane industry today is using many more simulations than actual wind tunnel tests, I think we will see a future where the researcher is sitting in front of a desktop and using simulations to predict the outcome of lots of “wet lab” experiments.
Instead of testing how a drug binds to a single protein in an experiment it will probably be possible to directly test how it binds to all different genetic mutations and what other molecules it could interact with. This will be a revolution, although I predict we will still use some “wet lab” experiments to confirm the final results.
Bio for Dr. Erik Lindahl
Erik Lindahl completed his undergraduate studies in engineering physics at Lund University, after which he received a Ph.D. in theoretical biophysics at the KTH Royal Institute of Technology in Stockholm in 2001. He has performed research at Groningen University, Stanford University, and the Pasteur Institute after which he assumed a position as assistant and later associate professor at Stockholm University in 2004. Since 2010 he has held dual appointments as professor of theoretical biophysics at KTH and professor of computational structural biology at Stockholm University. Since 2011 he has been a member of the Swedish Young Academy.
http://www.cbr.su.se (Center for Biomembrane Research)
http://www.scilifelab.se (Science for Life Laboratory)
http://www.e-science.se (Swedish e-Science Research Center)