Online, at conferences and in theory, manycore processors and the use of accelerators such as GPUs and FPGAs are being viewed as the next big revolution in high performance computing (HPC). If they can live up to the potential, these accelerators could someday transform how computational science is performed, providing much more computing power and energy efficiency.
And, in fact, they are already helping to drive significant scientific research projects — not bundled together in large systems, but rather one server at a time. In early December, a group of astronomers, physicists and HPC experts gathered at the SLAC National Accelerator Laboratory near San Francisco to discuss how GPUs and FPGAs are meeting their unique needs. The three-day workshop was co-organized by Lawrence Berkeley National Laboratory, NERSC, SLAC and Stanford’s Kavli Institute for Particle Astrophysics and Cosmology (KIPAC).
The workshop was organized as a part of an ongoing effort to develop infrastructure for enabling physics and astronomy data problems by utilizing these emerging technologies. More than a year ago under the leadership of Horst Simon of LBNL, John Shalf and Hemant Shukla also of LBNL with Rainer Spurzem of the Chinese Academy of Sciences agreed to establish a working collaboration. The workshop was held on a shoestring budget with help from Tom Abel of KIPAC.
“The participating scientific groups started with challenging problems that required parallel performance to meet real-time requirements,” said co-organizer Shukla. “The effective approach to solving such problems as wavefront sensing and real-time radio imaging is to identify the underlying algorithms for speedups and thereby solve common sets of problems.”
The problems shared a common issue — strong real-time constraints. One application is in solving the challenges in real-time control of adaptive optics systems for high-resolution, ground-based astronomy. The second was in radio telescope arrays in remote locations with only limited power. In the second case, the researchers needed the power of a highly parallel system, but a standard cluster computer on a rack would require more electricity than is available. Using GPU acceleration was just the ticket.
“Instead of starting with the technology and seeing if a problem could be solved, as is often the case, they had a problem and found the technology to solve it,” said co-organizer Shalf.
In both cases, the scientists needed a speedup in processor performance and discovered that new technologies such as GPUs and FPGAs provided the enhancements. Their needs were different than those of many other researchers, who look to HPC centers to run their applications on a larger number of processors rather than just running their applications faster.
“The current direction in supercomputing doesn’t address the needs of researchers who need to solve the same-size problem faster, as opposed to solving a bigger problem at the same speed,” Shalf said.
At the workshop, experts from Asia, Europe and North America got together to share information on solving problems in this area, as well as explore and discuss the scope and challenges of harnessing the full potential of these novel architectures for high performance computing. The workshop drew attendees from academia and industry in China, France, Germany, Japan, Taiwan and the United States. Future experiments such as the Large Synoptic Survey Telescope, Murchison Wide-filed Array, the next-generation SETI and the Allen Telescope Array participated in defining the future goals, as did industry leaders including NVIDIA, AMD, Apple, and Sun Microsystems.
Conference advisor Rainer Spurzem of the Chinese Academy of Sciences cited the “eclectic mix” of attendees as adding to the informative exchange of ideas and experience.
“Although the focus was on physics and astronomy applications, the solutions explored by the participants are likely to have broader impact across science and technology disciplines such as healthcare, energy, aerospace and others,” said workshop co-organizer Hemant Shukla of the Berkeley Lab Physics Division. “These emerging new techniques could lead to new systems and software that use both silicon and electrical power much more efficiently. As we move beyond today’s petascale systems, such efficiency is a necessity.”
Other groups are also meeting to explore how these emerging processor technologies can advance a broad range of scientific applications. The workshop at SLAC was held two weeks after the newly-formed Hybrid Multicore Consortium met for the first time at the SC09 (Supercomputing) conference in Portland, Ore. Co-founded by Berkeley Lab, Los Alamos and Oak Ridge national laboratories, the consortium seeks to address the challenge of re-engineering most of today’s scientific applications to take advantage of the resources provided by future hybrid multicore systems.
“While there is considerable excitement about the potential of multicore systems and harnessing their performance for computational science, reaching this goal will require a tremendous effort by both the application experts and software developers,” said LBNL’s Simon, one of three members of the consortium’s executive committee.
Afterward, Wei Ge of the Chinese Academy of Sciences wrote to the organizers, “It was a very informative and fruitful workshop and thank you very much again for your organization and kind invitation to us.”
And some participants were already looking ahead to future collaborations and building resources and communities.
“I have gained quite a bit of information and impressions throughout and I am in the process of transferring all that to our Sun community,” wrote Ferhat Hatay, who works in Strategic Engagements at Sun Microsystems. “We are most interested in contributing to collaboration efforts with the expertise, interest, and support from Sun as well as from our customer and user base.”