ceClub: Leveraging RDMA for Strongly Consistent Replicationat Large Scale
- Ken Birman (Cornell University)
- Wednesday, 18.1.2017, 14:30
- Taub 301
My work focuses on ways of replicating data in demanding settings, most recently the cloud. The cloud is a setting where copying information and replicating data or computation is common, yet it remains difficult to actually create new applications that leverage replication. Moreover, the existing libraries are very slow.
I’ll present Derecho, a blazingly fast C++ library for creating scalable data replication solutions. Derecho has a remarkably simple API and very low overheads, shifting as much work as possible from runtime to compile time. We separate control plane from data plane, sending data on RDMA over a multicast overlay, then using an RDMA-based distributed shared memory to coordinate delivery so as to ensure total ordering, data persistency or other desired properties such as virtually synchronous membership management. The overall framework is minimalist, yet can support complex applications, including ones that have subgroup structures or that shard data within larger sets of nodes. Derecho consistency subsumes both Paxos and virtual synchrony multicast into a single model.
Performance is very satisfying: As an atomic multicast, Derecho is tens of thousands of times faster than commonly used libraries, and as Paxos, it beats prior solutions by large factors while scaling far better than any existing Paxos solution.