Using RDMA for Efficient Index Replication in LSM Key-Value Stores

TL;DR

This paper introduces Talos, a novel RDMA-based approach for efficient index replication in LSM key-value stores, significantly reducing I/O and CPU overhead while maintaining high performance and scalability.

Contribution

Talos leverages RDMA and KV separation to enable efficient primary-backup replication with minimal I/O amplification and CPU overhead, avoiding costly backup compactions.

Findings

Reduces backup I/O amplification by up to 3x

Decreases CPU overhead in backups by up to 1.6x

Allows larger growth factors for space efficiency

Abstract

Log-Structured Merge tree (LSM tree) Key-Value (KV) stores have become a foundational layer in the storage stacks of datacenter and cloud services. Current approaches for achieving reliability and availability avoid replication at the KV store level and instead perform these operations at higher layers, e.g., the DB layer that runs on top of the KV store. The main reason is that past designs for replicated KV stores favor reducing network traffic and increasing I/O size. Therefore, they perform costly compactions to reorganize data in both the primary and backup nodes, which hurts overall system performance. In this paper, we design and implement Talos, an efficient rack-scale LSM-based KV store that aims to significantly reduce the I/O amplification and CPU overhead in backup nodes and make replication in the KV store practical. We rely on two observations: (a) the increased use of RDMA in the datacenter, which reduces CPU overhead for communication, and (b) the use of KV separation that is becoming prevalent in modern KV stores. We use a primary-backup replication scheme that performs compactions only on the primary nodes and sends the pre-built index to the backup nodes of the region, avoiding all compactions in backups. Our approach includes an efficient mechanism to deal with pointer translation across nodes in the region index. Our results show that Talos reduces in the backup nodes, I/O amplification by up to $3\times$, CPU overhead by up to $1.6\times$, and memory size needed for the write path by up to $2\times$, without increasing network bandwidth excessively, and by up to $1.3\times$. Overall, we show that our approach has benefits even when small KV pairs dominate in a workload (80%-90%). Finally, it enables KV stores to operate with larger growth factors (from 10 to 16) to reduce space amplification without sacrificing precious CPU cycles.