Data Sharing Analysis of Emerging Parallel Media Mining Workloads

Abstract

This paper characterizes the sharing behavior of emerging parallel media mining workloads for chip-multiprocessors. Media mining refers to techniques whereby users retrieve, organize, and manage media data. These applications are important in defining the design and performance decisions of future processors. We first show that the sharing behaviors of these workloads have a common pattern that the shared data footprint is small but the sharing activity is significant. Less than 15% of the cache space is shared, while 40% to 90% accesses are to the shared footprint in some workloads. Then, we show that for workloads with such significant sharing activity, a shared last-level cache is more attractive than private configurations. A shared 32MB last-level cache outperforms a private cache configuration by 20 – 60%. Finally, we show that in order to have good scalability on shared caches, thread-local storage should be minimized when building parallel media mining workloads.KeywordsFace DetectionCache SizeCache LineSharing BehaviorSoccer VideoThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.