|
Presented at: MultiCore Expo 2008, on April 1-3, 2008 at Santa Clara, CA.
Abstract: On traditional MP systems, poor scalability is frequently observed due to software design issues, with problems such as hot locks and data sharing common culprits. Dealing with these problems has required detailed knowledge of both the application and the target system, making this a cumbersome process. On multicore processors, because threads share a common L2 cache, these problems now have a much smaller impact on scalability, making the development of multithreaded applications significantly simpler. Further, the inter-thread communication latency is reduced by an order of magnitude or more. As a result, it is possible to thread even short-duration tasks, enabling more complete parallelization and a corresponding performance boost.In this presentation we illustrate the ability of multicore processors to not only simplify the threading of an application, but to also enable entirely new styles of parallel optimization that were not considered practical on traditional MP systems. We will demonstrate this concept, which is referred to as micro-parallelization, using a number of different examples that clearly illustrate the potential benefits to key application spaces.
Bio: Lawrence Spracklen is a senior staff engineer in the Architecture Technology Group at Sun Microelectronics, that is focused on architecting and modeling next-generation SPARC processors. Dr. Spracklen's main focus is improving application performance via both hardware and software techniques. Dr. Spracklen has over 10 publications, such as the article titled the “The coming wave of multithreaded chip multiprocessors” in the June 2007 edition of the International Journal of Parallel Programming. Dr. Spracklen has presented at IEEE and ACM conferences, acted as session chair and is a frequent member of program committees. Dr. Spracklen holds 3 US patents, and has almost twenty patents pending.
Presentation: “Multicore Processors and Microparallelism“
|