KPart: A Novel Technique for Partitioning Shared Caches

KPart is a tool for partitioning the last-level cache (LLC) dynamically between co-running applications on an Intel multicore processor with way-partitioning support. KPart is designed to sidestep a key limitation of way-partitioning: coarse-grained partition sizes.

KPart profiles the cache needs of co-running applications periodically, then groups the applications based on their cache profiles into clusters and maps cache partitions to these clusters. Applications within a cluster share their assigned partition, while applications across clusters remain isolated. More details on how profiling, clustering, and cache assignments are done are available in our HPCA'18 paper.