Quantifying behavioral differences between multimedia and general-purpose workloads

  • Authors:

  • Lieven Eeckhout;Koen De Bosschere

  • Affiliations:

  • Department of Electronics and Information Systems (ELIS), Ghent University, Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium;Department of Electronics and Information Systems (ELIS), Ghent University, Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium

  • Venue:
  • Journal of Systems Architecture: the EUROMICRO Journal
  • Year:
  • 2003

Quantified Score

Hi-index 0.00

Visualization

Abstract

Multimedia workloads are becoming increasingly more common on general-purpose computing systems. However, little quantitative results are available about the behavior of these workloads. This paper is a first step in quantifying and understanding the behavioral differences (if any) between multimedia and general-purpose workloads. This is done by comparing program characteristics of multimedia applications (coming from the MediaBench suite, the X benchmarks from the SimpleScalar distribution, plus a set of MPEG-4 like algorithms) and general-purpose applications (coming from the SPECint95 and the SPECint2000 benchmark suite). In addition to presenting a database of program characteristics, we conclude that (i) multimedia applications have less memory operations and less control operations in their instruction mix than general-purpose workloads, and thus are computationally intensive; (ii) multimedia and general-purpose applications exhibit comparable levels of instruction-level parallelism; (iii) multimedia applications also suffer from hard-to-predict branches and (iv) the instruction stream as well as the data stream of multimedia applications exhibit more spatial and more temporal locality than general-purpose applications. These results were obtained using statistical tests, namely the t-test and the Mann-Whitney test.