Colloquia, Seminars and Conference News
Title : Disk-Based Parallel Computing: A New Paradigm
Date : November 1, 2006. (10:30 am) Tea starts half an hour before each seminar
Location: ITEB 336
Speaker : Gene Cooperman
Abstract:
Today's computer architectures are more complex than in the past. To give a few cases: (1) CPU speeds are stagnating below the 4 GHz barrier, and we are compensating by moving toward multi-core CPUs; (2) RAM is becoming very cheap, but the motherboards don't have room for much more than the standard 4 GB RAM; (3) the pressure of those multi-core CPU chips on RAM continues to outstrip the available bandwidth of the bus to RAM. (4) disk capacities grow and keep up with the performance of multi-core CPUs (in a few months, a disk will have 1 terabyte), but the bandwidth of disk is hardly increasing at all. These disruptive times call for drastic measures. High performance computing requires not only higher CPU performance, but also ever larger storage. We must use disk as the new RAM. A modest 50-node dual-CPU cluster will soon have 50 local 1 terabyte disks. The bandwidth of 50 disks is about 2.5 GB/s, which is similar to the bandwidth of a RAM subsystem. Hence, we can treat the many local disks as a single extended RAM subsystem, _providing_ that we promise to use it only in streaming access, rather than in random access. Space-time tradeoffs allow us to treat disk as extended RAM, and to blow up our space requirements in order to reduce our time requirements. We provide some example algorithms for search and enumeration. Curiously, some of these revised algorithms are not limited by disk speed or CPU speed, but instead by RAM speed!
Bio:Gene Cooperman received his B.S. from the University of Michigan in 1974, in honors Mathematics and Physics. He received his Ph.D. in Applied Mathematics from Brown University in 1978. After a 2-year post-doc at Michigan State University, he joined GTE Laboratories for six years, where he became Principal Member of Technical Staff. He joined the College of Computer Science at Northeastern University in 1986, and became a full professor in 1992.
He has over 70 refereed publications, and has been awarded 15 grants from the National Science Foundation. He is the head of the High Performance Computing Lab at Northeastern University. He is also the director of the Institute for Complex Scientific Software, an inter-disciplinary collaboration across five departments at Northeastern University. He has graduated three Ph.D. students, and currently has 6~Ph.D. students. His research group has been developing the TOP-C (Task Oriented Parallel C/C++) Project for over ten years. Its goal is the easy parallelization of sequential software on commodity hardware. One notable milestone along this path was the first parallelization of Geant4, a million-line C++ program to simulate particle-matter interaction. Geant4 was developed by over a hundred physicists around the world, and is used to model experiments designed for the future large hadron collider being built at CERN in Geneva. The TOP-C parallelization is distributed with Geant4, and is also a part of the GAP interactive system (Groups, Algorithms and Programming), and soon GCL (GNU Common Lisp).
[Back]
