Record Detail

As the number of cores in chip multi-processor systems increases, the contention over shared last-level cache (LLC) resources increases, thus making LLC optimization critical, especially for embedded systems with strict area/energy/power constraints. We propose cache partitioning with partial sharing (CaPPS), which reduces LLC contention using cachepartitioningandimprovesutilizationwithsharingconfiguration.Sharingconfiguration enables the partitions to be privately allocated to a single core, partially shared with a subset ofcores,orfullysharedwithallcoresbasedontheco-executingapplications’requirements. CaPPS imposes low hardware overhead and affords an extensive design space to increase optimization potential. To facilitate fast design space exploration, we develop an analytical model to quickly estimate the miss rates of all CaPPS configurations using the applications’ isolated LLC access traces to predict runtime LLC contention. Experimental results demonstrate that the analytical model estimates cache miss rates with an average error of only 0.73% and with an average speedup of 3505× as compared to a cycle-accurate simulator. Due to CaPPS’s extensive design space, CaPPS can reduce the average LLC miss rate by as much as 25% as compared to baseline configurations and as much as 14–17% as compared to prior works.