Intelligent cache management for heterogeneous memory systems

Abstract

DRAM caches are important for enabling effective heterogeneous memory systems that can transparently provide the bandwidth of high-bandwidth memories and the capacity of high-capacity memories. This dissertation investigates enabling intelligent cache management for tag-inside-cacheline DRAM cache designs. Such a DRAM cache uses a direct-mapped design, co-locates the tag and data within the DRAM array, and streams out the tag and the data concurrently on an access. The direct-mapped design has been shown to be effective for enabling low latency and bandwidth-efficient tag access. However, such a direct-mapped design can have lower hit-rate and high bandwidth cost to confirm misses. This dissertation investigates simple architectural techniques to improve the hit-rate and bandwidth consumption of such DRAM caches by enabling associativity, replacement policies, and reduced miss probes at low bandwidth cost to improve the performance of heterogeneous memory systems.