MLPaware dynamic cache partitioning

Abstract. Dynamic partitioning of shared caches has been proposed to improve perfor-mance of traditional eviction policies in modern multithreaded architectures. All existing cache partitioning algorithms work on the number of misses caused by each thread and treat all misses equally. However, it has been shown that cache misses cause different impact in performance depending on the distribution of the Memory Level Parallelism (MLP) of the application L2 misses: clustered misses share their miss penalty as they can be served in parallel, while isolated misses have a greater impact as the memory latency is not shared with other misses. We take this fact into account and propose a new dynamic cache partitioning algorithm that considers misses differently depending on their influence in throughput. Our proposal obtains improvements over traditional traditional eviction policies up to 63.9 % (10.6 % on average) and it also outperforms previous dynamic partitioning proposals by up to 15.4% (4.1 % on average) in a four-core architecture. Finally, we have used a sampling technique to propose a practical implementation with a hardware cost under 1 % of the total L2 cache size.