MIT News Office series: Multicore Computing - a multisided EECS research effort

SHARE:
February 28, 2011

The MIT News Office has offered, starting February 23, 2011, a five part series of in-depth articles about MIT's Computer Science research titled "The Multicore Future." All articles are authored by News Office staff writer Larry Hardesty.

Part One of the series begins with an examination of current hardware 'dead ends' in terms of growth in chip speed and capacity and how the current development of multicore chips will necessitate the needs for communication between distributed processors and managing shared data. Since August 2010, the U.S. Department of Defense’s Defense Advanced Research Projects Agency announced creation of an $80 billion funded effort in “ubiquitous high-performance computing.” Members of the MIT Department of Electrical Engineering and Computer Science Department are now engaged in a major part of this effort. Project Angstrom, involving 19 MIT researchers (so far) and headed by EECS faculty member Anant Agarwal was created to investigate hardware issues of multicore processors. Read more of Part One, Feb. 23, 2011, "Designing the hardware."

Part Two of the Multicore series, titled "The Next Operating System" drills down to the ways that a multicore operating system will need to be more self-aware — to have better information about the computer’s performance as a whole — and to have more control of the operations executed by the hardware. Other EECS Department faculty members including Martin Rinard, Saman Amarasinghe, Frans Kaashoek and Nickolai Zeldovich and Srini Devadas (current Interim EECS Department Head) are looking, respectively, at ways to cut corners within processes (perforated loops), smart selection of the most efficient algorithms for specific tasks, cross-core cache referencing for more efficient operation (avoiding the system's main memory), limiting the number of operating-system subroutines that require that privileged access, and alternatively working directly with caches that already have the needed data. Read more of Part Two "The Next Operating System".

Part Three, titled "Retooling Algorithms" appearing Friday, Feb. 25, looks at how EECS faculty member Charles Leiserson and his team have been designing parallel algorithms since the 1990s -- in other words, using the technique known as divide-and-conquer to enable a computer to tailor an algorithm’s execution to available resources. Read more --especially about Leiserson's program developed for chess-playing that competed with IBM’s Deep Blue in 1995.

Part Four titled "No backtalk," looks at the way EECS research faculty are trying to minimize the amount of communication between cores to make parallel algorithms more efficient. From a discussion of the work by members of Leiserson's group on FFTs, fast Fourrier Transforms, one of the most frequently used classes of algorithms in computer science, used for signal processing, image processing, and data compression, among other things, to a description of the more recent work of EECS Professor Anantha Chandrakasan and (former) PhD student Vivienne Sze who have proposed a new, parallel version of an algorithm common to most computer video systems. Their work, now under review with the MPEG and ITU-T standards bodies, may well become part of the standard for video/image processing. Read more.

Part Five of the MIT News Office series on multicore computing is titled "Language barrier". It is subtitled "To take advantage of multicore chips, programmers will need software development systems that let them express themselves in fundamentally new ways." The problem is posed: "For decades, computer scientists tried to develop software that could automatically turn a conventional computer program — a long sequence of instructions intended to be executed in order — into a parallel program — multiple sets of instructions that can be executed at the same time. Now, most agree that that was a forlorn hope: Code that can be parallelized is too hard to recognize, and the means for parallelizing it are too diverse and context-dependent."

The discussion which follows in this final part of the Multicore series includes the work of CSAIL and EECS faculty members Charles Leiserson and Saman Amarasinghe and former CSAIL PhD graduate and postdoc Alexey Radul. Read more.