Using Locality and Interleaving Information to Improve Shared Cache Performance
Publication or External Link
The cache interference is found to play a critical role in optimizing cache allocation among concurrent threads for shared cache. Conventional LRU policy usually works well for low interference workloads, while high cache interference among threads demands explicit allocation regulation, such as cache partitioning. Cache interference is shown to be tied to inter-thread memory reference interleaving granularity: high interference is caused by ne-grain interleaving while low interference is caused coarse-grain interleaving. Proling of real multi-program workloads shows that cache set mapping and temporal phase result in the variation of interleaving granularity. When memory references from dierent threads map to disjoint cache sets, or they occur in distinct time windows, they tend to cause little interference due to coarse-grain interleaving. The interleaving granularity measured by runlength in workloads is found to correlate with the preference of cache management policy: ne-grain interleaving workloads perform better with cache partitioning, and coarse-grain interleaving workloads perform better with LRU.
Most existing shared cache management techniques are based on working set locality analysis. This dissertation studies the shared cache performance by taking both locality and interleaving information into consideration. Oracle algorithm which provides theoretical best performance is investigated to provide insight into how to design a better practical policy. Proling and analysis of Oracle algorithm lead to the proposal of probabilistic replacement (PR), a novel cache allocation policy.
With aggressor threads information learned on-line, PR evicts the bad locality blocks of aggressor threads probabilistically while preserving good locality blocks of non-aggressor threads. PR is shown to be able to adapt to the different interleaving granularities in different sets over time. Its flexibility in tuning eviction probability also improves fairness among thread performance. Evaluation indicates that PR outperforms LRU, UCP, and ideal cache partitioning at moderate hardware cost.
For single program cache management, this dissertation also proposes a novel technique: reuse distance last touch predictor (RD-LTP). RD-LTP is able to capture reuse distance information, which represents the intrinsic memory reference pattern.
Based on this improved LT predictor, an MRU LT eviction policy is developed to select the right victim at the presence of incorrect LT prediction. In addition to LT predictor, another predictor: reuse distance predictors (RDPs) is proposed, which is able to predict actual reuse distance values. Compared to various existing cache management techniques, these two novel predictors deliver higher cache performance with higher prediction coverage and accuracy at moderate hardware cost.