An In-Memory Key-Value Caching System Based On Hybrid DRAM/NVM Memory

In-memory key-value caching systems are widely deployed in data centers.However,Dynamic Random Access Memory(DRAM)can't meet increasing memory requirements of in-memory key-value caching systems for its technology and energy limitations.Emerging Non-Volatile Memory(NVM)technologies promise much higher memory density and energy efficiency than DRAM,at the expense of higher read/write latency and higher write energy.Hybrid memory systems composed of DRAM and NVM have the potential to provide very large capacity of memory while maintain high performance.However,there remains challenges to deploy DRAM-based in-memory key-value caching systems in hybrid memory systems.The performance and energy efficiency of in-memory key-value caching systems on hybrid memory systems have not been fully explored yetIn this paper,we propose HMCached,an in-memory key-value caching system built on a hybrid DRAM/NVM system.HMCached utilizes an application-level object access counting mechanism to identify frequently-accessed objects in NVM,and migrates them to fast DRAM to reduce the costly NVM accesses.Compared to previous work,HMCached avoids any modifications to hardware and OS.Meanwhile,the object access counting cause negligible runtime overhead.We propose an NVM-friendly index structure.By storing the frequently-updated portion of object metadata in DRAM solely,HMCached can further mitigate the NVM accesses.We choose slab allocation as the memory allocation mechanism for HMCached.For the slab calcification problem,we propose a slab reassignment policy to solve the calcification of DRAM.We evaluate HMCached with Zipfian-like workloads.Experiment results show that HMCached significantly reduces NVM accesses by about 70% compared to the vanilla Memcached,and improves application performance by up to 56%.Moreover,compared to a DRAM-only system,HMCached achieves 86% of performance and 47% reduction of energy consumption for realistic(read-intensive)workloads while significantly reducing the DRAM usage by 80%.