Title: Computational force and computational energy: Two new approaches to performance analysis Speaker: Robert Numrich, Minnesota Supercomputing Institute, University of Minnesota, Minneapolis, MN Date/Time: Monday, June 2, 2008, 3:00 – 4:30 pm MST Location: CSRI Building, Room 90 (Sandia NM) Brief Abstract: Dimensional analysis applied to timing formulas for numerical algorithms yields self-similarity surfaces that describe how an algorithm behaves in terms of dimensionless ratios of computational forces. Computational forces are also known as computational intensities, but identifying them as forces allows us to analyze performance in terms of the contention between hardware forces and software forces. Two machines may have different values for latency, bandwidth and computational power, but if the forces defined by these quantities are the same, the two machines scale the same way along the same path on the same self-similarity surface. Computational energy allows us to analyze performance at the instruction level as the program executes. From the issue time and completion time for each instruction, we can calculate the computational kinetic energy as a function of time and plot a computational energy spectrum of the program from beginning to end. Sharp spikes in the spectrum correspond to important events during execution such as cache misses. I will show an example obtained by simulation of a small program using the Apple G5 simulator...CSRI POC: Mike Heroux, (320) 845-7695 |