|
September 15, 2008 Article #20540 Volume 127, Issue 3 HPC
|
Anton is in the works: a massively parallel supercomputer
Grand Challenge: Millisecond-scale Molecular Dynamics Simulations
Anton May Be the Answer
|
|
In his paper "Grand Challenge: Millisecond-scale Molecular Dynamics Simulations," Eduardo D'Azevedo, a member of the Computer Science and mathematics Division at Oak Ridge National Laboratory, discusses the work in molecular dynamics (MD) simulation being conducted by D.E. Shaw Research (DESRES), a company headed by David E. Shaw, a senior research fellow at the Center for Computational Biology and Bioinformatics, Columbia University.
Molecular dynamics simulation, D'Azevedo explains, is an important tool for the modeling of protein-size systems (25,000–50,000 atoms in water) in applications involving such things as the development of new drugs. The grand challenge for MD researchers developing the ability to simulate such systems as long trajectories, in the millisecond time scale, where biologically interesting phenomena occur.
Among these phenomena are the folding of proteins, the binding of drugs to molecular targets, interactions between proteins, and the dynamics of conformational changes in macromolecules, according to D'Azevedo.
By way of making this challenge readily understandable, the author writes that a single processor can simulate about one nanosecond in a day, while a massively parallel code might be able to simulate about one hundred nanoseconds per day. Meeting this grand challenge will thus require close to a hundred-fold speedup, which in turn will require new massively parallel architectures and innovative algorithms, he posits. Clearly a job for HPC.
Researchers at DESRES have developed a new MD code, called Desmond, that uses the novel "new territory" or NT parallel algorithms and numerical techniques to achieve very high performance levels on hardware ranging from a conventional commodity cluster to high-performance machines, D'Azevedo writes, adding that a version of Desmond will be available at no cost for non-commercial use at universities and other not-for-profit re-search institutions. A commercial version of Desmond is expected to be available later this year.
Yet another project at DESRES that the author notes is the development and implementation of Anton, a specialized, massively parallel supercomputer that was designed to execute MD simulations as many as hundreds of times faster than had been possible previously.
The author explains that Anton achieves these speeds by employing “arithmetic specialization," which provides flexibility or programmability only where needed; elsewhere, hardware is tailored for high speeds.
D'Azevedo writes that one segment of Anton consists of 512 application-specific integrated circuits (ASICs), connected as an 8 × 8 × 8 three-dimensional torus. This 3D torus reflects the physical space being simulated, with its nearest-neighbor connections and periodic boundary conditions.
Communication on Anton is timed so that data flows to a processor only when needed, thus reducing communication overhead is reduced, as is the number of off-chip memory accesses. By these means, each ASIC achieves about 500 times the performance of a general-purpose microprocessor while using the same amount of power.
Delving further into the anatomy of Anton, D'Azevedo notes that each of Anton’s chips contains 32 particle-interaction pipelines—specialized fixed hardware that can efficiently compute the energy, potential, and force vectors.
In dealing with algorithms that are less regular or more likely to change, such as as calculations of bonded interactions, bond-length constraints, and other alternative integration techniques, Anton uses more flexible programmable hardware. This flexible subsystem exploits three forms of parallelism: multicore parallelism, with four Tensilica cores (custom instructions) and eight geometry cores (instruction-level parallelism), and single instruction, multiple data (SIMD) parallelism, with 3D vectors manipulated as a single operation, according to the author. Each geometry core is implemented as a very long instruction word (VLIW) processor.
Test runs on a real (prototype) Anton processor have produced very promising
performance results. By the end of 2008, when Anton is expected to be completed, it will be possible to execute regular millisecond-scale MD simulations, which should yield some exciting new discoveries, D'Azevedo concludes.
Send to a Colleague
Printer-Friendly Version
Grand Challenge: Millisecond-scale Molecular Dynamics Simulations
|
|
 
 
 
 
 
 
|
|
|
Customized news reports about Sun Microsystems. Just the news you need, none of what you don't.
50,000+ Members. 20,000+ Articles Published since 1998. |
Copyright © 1998-2008 System News, Inc.