Using Persistent Memory and RDMA for Ceph Client Write-back Caching | SNIA

Abstract

There are many reasons Ceph RBD is a good choice for providing block storage to virtual machines, but RBD’s write tail latency (often 10’s of mS) is not always acceptable. In this session we describe Intel’s work in the Ceph community to add a replicated write-back cache to Ceph librbd. The Replicated Write Log (RWL) uses persistent memory in the client node to significantly improve write tail latency. RWL can use the replication capability of the open-source persistent memory APIs (PMDK) to mirror the write back cache over RDMA to a Ceph OSD node. Once initialized, zero CPU cycles are consumed on the replica node unless the client fails.

Learning Objectives:
1. RWL design overview, and how it fits into Ceph
2. Prototype performance measurement results
3. MySQL benchmark on RBD with and without prototype RWL