7.1.2008: Guest lecture by Prof. Alan Jay Smith, UC Berkeley, HIIT Spektri 1:30-4:00 PM

Prof. Alan Jay Smith from University of California at Berkeley will give a guest lecture in HIIT/Spektri on Monday, 7th 2008, 1:30 pm - 4 pm. All interested parties are welcome! Title and abstract follow.

Multimedia Workloads, Architectures to Process Them, Their Performance and the Use of Caches for Multimedia Workloads

Alan Jay Smith
CS Division
UC Berkeley
Berkeley, California

Monday, Jan 7th 2008, 1:30 pm - 4 pm
HIIT, Spektri Pilotti, 3rd floor, Metsänneidonkuja 4, Espoo

The last decade has seen the integration of audio, video, and 3D graphics into existing workloads as well as the emergence of new workloads dominated by the processing of these forms of media. Unfortunately, widely accepted benchmarks which capture these new workloads in a realistic way have not emerged. We present the Berkeley multimedia workload, which was developed to facilitate our own studies on architectural support for multimedia.

The caching behavior of multimedia applications has been described as having high instruction reference locality within small loops, very large working sets, and poor data cache performance due to non-locality of data references. Despite this, there is no published research deriving or measuring these qualities. Utilizing the previously developed Berkeley Multimedia Workload, we present the results of execution driven cache simulations with the goal of aiding future media processing architecture design. Our analysis examines the differences between multimedia and traditional applications in cache behavior.  We find that multimedia applications actually exhibit lower instruction miss ratios and comparable data miss ratios when contrasted with other widely studied workloads.  In addition, we find that longer data cache line sizes than are currently used would benefit multimedia processing.

Many microprocessor instruction sets include instructions for accelerating multimedia applications such as DVD playback, speech recognition and 3D graphics.  Despite general agreement on the need to support this emerging workload, there are considerable differences between the instruction sets that have been designed to do so.  We present a study of the performance of five instruction sets on kernels extracted from a broad multimedia workload. Each kernel was recoded in the assembly language of the five multimedia extensions.  We compare the performance of each extension against other architectures as well as to the original compiled C performance. From our analysis we determine how well multimedia workloads map to current architectures, what was useful and what was not. We also propose two enhancements to current architectures: strided memory operations, and superwide registers.

The work to be presented was done with Nathan Slingerland, formerly a graduate student at UC Berkeley.

Last updated on 22 Apr 2008 by Teemu Mäntylä - Page created on 7 Jan 2008 by Martti Mäntylä