Fast in-memory storage systems : two aspects

Abstract

This dissertation addresses two challenges relating to in-memory storage systems. The first challenge is storing and retrieving data at a rate close to the capabilities of the underlying memory system, particularly in the face of parallel accesses from multiple cores. We present Masstree, a high performance in-memory key-value store that runs on a single multi-core server. Masstree is derived from a concurrent B+tree. It provides lock-free reads for good multi-core performance, which requires special care to avoid writes interfering with concurrent reads. To reduce time spent waiting for memory for workloads with long common key prefixes, Masstree arranges a set of B+trees into a Trie. Masstree uses software prefetch to further hide DRAM latency. Several optimizations improve concurrency. Masstree achieves millions of queries per second on a 16-core server, which is more than 30x as fast as MongoDB [6] or VoltDB [17]. The second challenge is replicating storage for fault-tolerance without being limited by slow writes to stable disk storage. Lazy VSR is a quorum-based replication protocol that is fast and can recover from simultaneous crashes of all the replicas as long as a majority revive with intact disks. The main idea is to acknowledge requests after recording them in memory, and to write updates to disk in the background, allowing large batched writes and thus good performance. A simultaneous crash of all replicas may leave the replicas with significantly different on-disk states; much of the design of Lazy VSR is concerned with reconciling these states efficiently during recovery. Lazy VSR's client-visible semantics are unusual in that the service may discard recent acknowledged updates if a majority of replicas crash. To demonstrate that clients can nevertheless make good use of Lazy VSR, we built a file system backend on it. Evaluation shows that Lazy VSR achieves much better performance than a version of itself with traditional group commit. Lazy VSR achieves 1.7 x the performance of ZooKeeper [42] and 3.6 x the performance of MongoDB [6].