p(2)KVS: a Portable 2-Dimensional Parallelizing Framework to Improve Scalability of Key-value Stores on SSDs

Attempts to improve the performance of key-value stores (KVS) by replacing the slow Hard Disk Drives (HDDs) with much faster Solid-State Drives (SSDs) have consistently fallen short of the performance gains implied by the large speed gap between SSDs and HDDs, especially for small KV items. We experimentally and holistically explore the root causes of performance inefficiency of existing LSM-tree based KVSs running on powerful modern hardware with multicore processors and fast SSDs. Our findings reveal that the global write-ahead-logging (WAL) and index-updating (MemTable) can become bottlenecks that are as fundamental and severe as the commonly known LSM-tree compaction bottleneck, under both the single-threaded and multi-threaded execution environments. To fully exploit the performance potentials of full-fledged KVS and the underlying high-performance hardware, we propose a portable 2-dimensional KVS parallelizing framework, referred to as p(2)KVS. In the horizontal inter-KVS-instance dimension, p(2)KVS partitions a global KV space into a set of independent subspaces, each of which is maintained by an LSM-tree instance and a dedicated worker thread pinned to a dedicated core, thus eliminating structural competition on shared data structures. In the vertical intra-KVS-instance dimension, p(2)KVS separates user threads from KVS-workers and presents a runtime queue-based opportunistic batch mechanism on each worker, thus boosting process efficiency. Since p(2)KVS is designed and implemented as a user-space request scheduler, viewing WAL, MemTables, and LSM-trees as black boxes, it is nonintrusive and highly portable. Under micro and macro-benchmarks, p(2)KVS is shown to gain up to 4.6x write and 5.4x read speedups over the state-of-the-art RocksDB.