Separated Allocator Metadata in Disaggregated In-Memory Databases: Friend or Foe?

Abstract

Memory allocation has a significant impact on the performance of in-memory databases. While state-of-the-art memory allocators work well in DRAM-only setups, some of their design decisions might no longer yield efficiency if data is tiered to disaggregated memory or secondary memory tiers. In this work, we study the performance impact of metadata in memory allocators and their tiering to disaggregated memory in the context of in-memory databases for the first time. We show how to separate metadata and application data by the example of jemalloc, which is widely used for data-intensive applications, and study performance effects for different workloads.