One Supercomputer Architecture Will Integrate Multiple Processing Technologies
Cray Inc. today announced plans to develop supercomputers that will take the concept of heterogeneous computing to an entirely new level by integrating a range of processing technologies in a single platform. These "adaptive supercomputing" systems will be able to solve scientific and engineering problems more quickly -- and make programmers and end users more productive -- by adapting processing to the requirements of each application.
Over the next few years, Cray supercomputers will combine standard microprocessors (scalar processing), vector processing, multithreading and hardware accelerators in one high-performance computing (HPC) platform that uses the industry-standard Linux operating system. Powerful compilers and other software will automatically match an application to the processor blade that is best suited for it.
The adaptive supercomputing approach means that users will not have to bias their HPC platform decisions toward the most widely used applications. It also means users will not have to make costly and time-consuming programming alterations so their applications will run efficiently on a particular platform. Cray enjoys a head start in adaptive supercomputing because the company's depth of expertise and intellectual property encompasses the various processing types and their related compilers.
"With adaptive supercomputing, Cray is once again providing the HPC market with new, innovative approaches," said Dr. Earl Joseph, research vice president of high-performance systems at IDC. "By targeting HPC systems with a broad range of processor choices and innovative software, Cray is poised to reverse the current practice of altering applications to fit platforms. Recent IDC end-user surveys show that HPC users place a high value on innovative approaches that enable broad usability, especially those that have the possibility of taking computing to a new level, or make large-scale systems easier to use."
"Cray's adaptive supercomputing plans are genuinely exciting," said Robert Harrison, an Oak Ridge National Laboratory researcher and winner of the IEEE Computer Society's Sidney Fernbach Award. "Computational chemistry cuts across many disciplines, from materials science to biology and environmental science. A single computing resource that can solve multidisciplinary and multiscale problems, while also achieving high levels of performance and scalability, will be extremely valuable in our field. It will make it easier to develop and deploy new science, because we will be able to focus more on the science and worry less about the computers."
Future Importance for HPC
There are two fundamental reasons why adaptive supercomputing will be important to the future of HPC. First, applications in areas ranging from aircraft design to climate modeling have become much more complex. In fact, many are really collections of separate sub-applications with varied characteristics and processing requirements.
Second, recent advancements in multicore microprocessors that have resulted in more tightly coupled processing have boosted performance, but still provide just one type of processing. Furthermore, multicore processors create additional challenges in scalable systems due to limitations in bandwidth, memory access and software scalability.
"Different applications run best on different types of processors, but high-performance computers typically offer only one type of processor," said Cray Chief Technology Officer Steve Scott. "Even today's heterogeneous computing environments really just loosely link differently architected computers, rather than offering true processing heterogeneity and adaptability. Cray will build supercomputers that can adapt to the applications, instead of forcing the applications to adapt to the supercomputers. Over time, these systems will include intelligence that can examine an application, determine which processing technique will work best with it and then handle the application accordingly -- without user intervention."
Adaptive Supercomputing Roadmap
Cray will implement adaptive supercomputing in phases. The first phase, code-named "Rainier," will create an integrated user environment across all of Cray's platforms. The second phase will result in a fully integrated multiarchitecture system, and the final phase will see the development of Cray systems that incorporate dynamic resource allocation using software that automates adaptive supercomputing.
The foundation for Cray's adaptive supercomputers will be a scalable infrastructure that employs the company's long-standing approach to high bandwidth and low latency. Processing technologies will leverage Cray's portfolio of expertise in vector and multithreading, as well as the AMD Opteron(TM) scalar processor technology that is currently part of the industry-leading Cray XT3(TM) and Cray XD1(TM) systems.
In November 2005, Cray and AMD announced an agreement that extended their successful relationship through the end of the decade. The two firms are also actively collaborating on Cray's Phase 3 proposal for the Defense Advanced Research Projects Agency (DARPA) High Productivity Computing Systems program, which has challenged HPC vendors to develop a new generation of supercomputers that can perform quadrillions of operations per second.