Performance Degradation in the Presence of Subnormal Floating-Point Values
Authors:
Orion Lawlor, Hari Govind, Isaac Dooley, Michael Breitenfeld, and Laxmikant Kale
Parallel Programming Laboratory, Department of Computer Science, University
of Illinois at Urbana-Champaign
In OSIHPA Workshop, at PACT05, September 2005.
Operating system interference in parallel programs can cause tremendous performance degradation. This paper discusses the interference caused by the quiet generation of subnormal floating point values. We analyze the performance impact of subnormal values in a parallel simulation of a stress wave propagating through a three dimensional bar. The floating-point exception handling mechanisms of various parallel architectures and operating systems lead to widely differing performance for the same program. We show that a parallel program will exhibit greatly amplified performance degredation due to this interference. In addition we provide an simple example program that demonstrates underflow on a single processor. Finally we suggest a novel option for fixing these undesired slowdowns.