Balance in machine architecture [electronic resource] : Bandwidth on board and offboard, integer/control speed and flops versus memory
- Washington, D.C. : United States. Dept. of Energy, 1992.
Oak Ridge, Tenn. : Distributed by the Office of Scientific and Technical Information, U.S. Dept. of Energy.
- Physical Description:
- Pages: (45 pages) : digital, PDF file
- Additional Creators:
- Fermi National Accelerator Laboratory, United States. Department of Energy, and United States. Department of Energy. Office of Scientific and Technical Information
- Restrictions on Access:
- Free-to-read Unrestricted online access
- The issues to be addressed here are those of balance'' in machine architecture. By this, we mean how much emphasis must be placed on various aspects of the system to maximize its usefulness for physics. There are three components that contribute to the utility of a system: How the machine can be used, how big a problem can be attacked, and what the effective capabilities (power) of the hardware are like. The effective power issue is a matter of evaluating the impact of design decisions trading off architectural features such as memory bandwidth and interprocessor communication capabilities. What is studied is the effect these machine parameters have on how quickly the system can solve desired problems. There is a reasonable method for studying this: One selects a few representative algorithms and computes the impact of changing memory bandwidths, and so forth. The only room for controversy here is in the selection of representative problems. The issue of how big a problem can be attacked boils down to a balance of memory size versus power. Although this is a balance issue it is very different than the effective power situation, because no firm answer can be given at this time. The power to memory ratio is highly problem dependent, and optimizing it requires several pieces of physics input, including: how big a lattice is needed for interesting results; what sort of algorithms are best to use; and how many sweeps are needed to get valid results. We seem to be at the threshold of learning things about these issues, but for now, the memory size issue will necessarily be addressed in terms of best guesses, rules of thumb, and researchers' opinions.
- Report Numbers:
- E 1.99:fnal/c-92/125
E 1.99: conf-9001172--1
- Other Subject(s):
- Published through SciTech Connect.
Terraflop workshop, Tallahassee, FL (United States), Jan 1990.
- Funding Information:
View MARC record | catkey: 14357440