TrickleKV: A High-Performance Key-Value Store on Disaggregated Storage With Low Network Traffic

Abstract

Disaggregated storage (DS) based on remote direct memory access (RDMA) network decouples compute and storage resources, thereby significantly improving resource utilization. While building key-value (KV) stores on DS benefits from these merits, existing fast KV stores suffer from network bandwidth contention and high latency under DS due to the non-negligible network amplification and high-overhead I/O stack. In this paper, we propose TrickleKV, a high-performance persistent KV store designed for DS. TrickleKV reduces network amplification and latency in three approaches: 1) TrickleKV proposes an efficient storage-side data filtering mechanism and a two-level cache structure with different granularities to reduce network traffic in the read process. 2) TrickleKV presents an efficient write buffer structure that includes asynchronous flushing and queue scheduling mechanisms to reduce network traffic in the write process. 3) TrickleKV designs a read-write decoupled user-space I/O stack and lightweight storage space management to reduce access latency. Evaluation results show that TrickleKV achieves $1.2\times $ – $7\times $ higher throughput and 30%- $7.4\times $ lower latency compared to state-of-the-art KV stores under DS.

Keywords

• Disaggregated storage
• key-value store
• NVMe over fabrics
• remote direct memory access