Reducing Waste in Memory Hierarchies

Abstract

Memory hierarchies play an important role in microarchitectural design to bridge the performance gap between modern microprocessors and main memory. However, memory hierarchies are inefficient due to storing waste. This dissertation quantifies two types of waste, dead blocks and data redundancy. This dissertation studies waste in diverse memory hierarchies and proposes techniques to reduce waste to improve performance with limited overhead.

This dissertation observes that waste of dead blocks in an inclusive last level cache consists of two kinds of blocks: blocks that are highly accessed in core caches and blocks that have low temporal locality in both core caches and the last-level cache. Blindly replacing all dead blocks in an inclusive last level cache may degrade performance. This dissertation proposes temporal-based multilevel correlating cache replacement to improve performance of inclusive cache hierarchies.

This dissertation observes that waste exists in private caches of graphics processing units (GPUs) as zero-reuse blocks. This dissertation defines zero-reuse blocks as blocks that are dead after being inserted into caches. This dissertation proposes adaptive GPU cache bypassing technique to improve performance as well as reducing power consumption by dynamically bypassing zero-reuse blocks.

This dissertation exploits waste of data redundancy at the block-level granularity and finds that conventional cache design wastes capacity because it stores duplicate data. This dissertation quantifies the percentage of data duplication and analyze causes. This dissertation proposes a practical cache deduplication technique to increase the effectiveness of the cache with limited area and power consumption.