Region-level Tracking for Scalable Directory Cache

dc.contributor.authorYoon, Hongil
dc.contributor.authorSohi, Gurindar S.
dc.date.accessioned2015-09-24T21:43:53Z
dc.date.available2015-09-24T21:43:53Z
dc.date.issued2015-04-19
dc.description.abstractTraditional coherence directories track sharing information at a cache-line granularity. In practice, however, as data sharing occurs at a coarser granularity in a large region of memory, common sharing patterns tend to be observed across multiple proximate lines. Hence, the directory entries for the lines replicate the same sharing information, resulting in inefficient use of space, power, and energy. In this paper, we empirically demonstrate "region-level sharing pattern locality", that is, a small number of distinct sharing patterns are observed across proximate lines within a large region of memory, e.g., a page unit. We leverage this phenomenon to propose a new representation of sharing information, called Region-level Sharing information Tracking (RST), that dynamically maintains common sharing information in a space-efficient manner at a region-level. Our experimental results based on conventional parallel and server workloads show that RST reduces over 75% of the area (and hence energy) compared to conventional directory caches, with almost negligible performance overhead.en
dc.identifier.citationTR1823en
dc.identifier.urihttp://digital.library.wisc.edu/1793/73603
dc.subjectdirectory cachesen
dc.subjectDirectory-based coherence protocolsen
dc.subjectcache coherenceen
dc.subjectregion-level sharing pattern localityen
dc.subjectRegion-level Sharing information Trackingen
dc.subjectRSTen
dc.subjectmulticoreen
dc.titleRegion-level Tracking for Scalable Directory Cacheen
dc.typeTechnical Reporten

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
TR1823.pdf
Size:
1.01 MB
Format:
Adobe Portable Document Format
Description:
tech report

License bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
license.txt
Size:
2.03 KB
Format:
Item-specific license agreed upon to submission
Description: