Data sharing analysis of emerging parallel media mining workloads

Authors:
Yu Chen;Wenlong Li;Junmin Lin;Aamer Jaleel;Zhizhong Tang
Affiliations:
Tsinghua National Laboratory for Information Science and Technology, Department of Computer Science and Technology, Tsinghua University, Beijing, China;Intel China Research Center, Beijing, China;Tsinghua National Laboratory for Information Science and Technology, Department of Computer Science and Technology, Tsinghua University, Beijing, China;Intel Corporation, Hudson, MA;Tsinghua National Laboratory for Information Science and Technology, Department of Computer Science and Technology, Tsinghua University, Beijing, China
Venue:
HiPC'08 Proceedings of the 15th international conference on High performance computing
Year:
2008

Citing 5
Cited 0

Real-time goal-mouth detection in MPEG soccer video

MULTIMEDIA '03 Proceedings of the eleventh ACM international conference on Multimedia
Towards the Parallelization of Shot Detection - a Typical Video Mining Application Study

ICPP '06 Proceedings of the 2006 International Conference on Parallel Processing
IEEE International Symposium on Workload Characterization

IISWC '07 Proceedings of the 2007 IEEE 10th International Symposium on Workload Characterization
A three-level scheme for real-time ball tracking

MCAM'07 Proceedings of the 2007 international conference on Multimedia content analysis and mining
Robust head tracking with particles based on multiple cues fusion

ECCV'06 Proceedings of the 2006 international conference on Computer Vision in Human-Computer Interaction

Quantified Score

Hi-index	0.00

Visualization

Abstract

This paper characterizes the sharing behavior of emerging parallelmedia mining workloads for chip-multiprocessors. Media mining refers to techniqueswhereby users retrieve, organize, and manage media data. These applicationsare important in defining the design and performance decisions of futureprocessors. We first show that the sharing behaviors of these workloads have acommon pattern that the shared data footprint is small but the sharing activity issignificant. Less than 15% of the cache space is shared, while 40% to 90% accessesare to the shared footprint in some workloads. Then, we show that forworkloads with such significant sharing activity, a shared last-level cache is moreattractive than private configurations. A shared 32MB last-level cache outperformsa private cache configuration by 20-60%. Finally, we show that in orderto have good scalability on shared caches, thread-local storage should be minimizedwhen building parallel media mining workloads.