LSBench
| Workload Characterization and Analysis of a Life Sciences Benchmark |
| Meghana Deodhar1, Tyler Olsen1,2, Raj Panda3, and Lizy John1 |
| 1Department of Electrical and Computer Engineering, The University of Texas at Austin |
| 2Freescale Semiconductor |
| 3IBM Systems and Technology Group, Austin, TX |
Abstract
Bioinformatics is a rapidly growing segment of high performance computing and an emerging technology. However, no industry standard benchmark exists for this area and very little research has been carried out for bioinformatics workloads. LSBench is a life sciences benchmark suite compiled by IBM's Systems and Technology Group, which consists of four benchmarks that were formed to be representative of a typical bioinformatics computation scenario. In this paper we perform a workload characterization of this suite across three architectures.
Our primary aim in this paper is to characterize each benchmark with respect to its instruction mix, cache behavior, branch predictability and other metrics across multiple architectures that will give us insight into these workloads. Furthermore, we perform a function-level analysis of program cache behavior and generate phase plots to visualize the life cycles of these benchmarks. Our major conclusions are that the cache miss rates show a noticeable difference among the different architectures and all four benchmarks have very distinct phase behavior. We found that phases are typically few in number and intervals belonging to the same phase show very little variation.
Download
The benchmark suite is available for download. The file is large (810 MB) and may take time to download. Use bzip2 and tar to uncompress.