Quantifying the Interference Caused by Subnormal Floating-Point Values

Ordinary floating point operations are slow when subnormal(also called denormalized or unnormalized) floating point values are used or produced. This paper describes how to quantify the extent of the slowdown caused by subnormal values by providing a simple micro-benchmark. Previous work has shown that the slowdown caused by subnormals is significant for some real world applications and that the occurrence of subnormals should not just be treated as an exceptional case. This paper both proposes a micro-benchmark and then analyzes the results for running the proposed micro-benchmark in both C and Java on a variety of modern micro-architectures. The results presented show that all common modern micro-architectures are susceptible to this type of undesirable slowdown.

Isaac Dooley and Laxmikant Kale
Quantifying the Interference Caused by Subnormal Floating-Point Values
In OSIHPA Workshop, at PACT06, September 2006

@INPROCEEDINGS{SubnormalOSIHPA06,
  author = "Isaac Dooley and Laxmikant Kale",
  title = "Quantifying the Interference Caused by Subnormal Floating-Point Values",
  booktitle = "Proceedings of the Workshop on Operating System Interference in High Performance Applications",
  year = "2006",
  month = "September",
  day = "17"
}