Caching Strategies for More Efficient Generational Garbage Collection

Abstract

Until the advent of generational garbage collection, page faults caused by garbage collection were a major source of bottleneck. Generational garbage collectors, by collecting smaller regions of the heap called generations, reduce the memory footprint of the collector and therefore the number of page faults caused by it. With page faults out of the way, the focus now is on cache misses due to garbage collection. The gap between the processor and memory cycle time is widening each year. Projections indicate that this increase is likely to continue. This makes cache performance an attractive area to study in order to tune the performance of a program.

In one such study, a strategy has been proposed to improve cache performance of generational garbage collectors by pinning the youngest generation in the L2 cache. In this thesis, we study an anomaly associated with this strategy, and propose a new realization of the pinning strategy that removes this anomaly, thereby making it more attractive. We apply the idea of an SNT (selectively non-temporal) cache to garbage collection. This helps reduce cache pollution and conflict misses in a direct mapped cache due to non-temporal accesses during garbage collection.

Simulation results show an average miss-rate reduction of 10% for 16 KB and 32 KB direct mapped L1 caches with SNT support. The improvement is greater in benchmarks with a large amount of live data.