HeterMM: applying in-DRAM index to heterogeneous memory-based key-value stores

Abstract

We propose HeterMM, a versatile framework that leverages in-DRAM indexes in KV stores on heterogeneous memory. HeterMM incorporates a plug-in programming model, allowing for the integration of various types of indexes. By prioritizing the maintenance of both index and hot data in DRAM, HeterMM maximizes the utilization of the superior performance of DRAM. Our evaluation demonstrates that HeterMM outperforms existing state-of-the-art frameworks that convert in-DRAM indexes to persistent ones. Furthermore, HeterMM can surpass NVM-specific KV stores by carefully selecting the appropriate index for specific scenarios.