Leveraging Emerging Data Center Technologies to Build High-Performance Data Stores

Abstract

Distributed in-memory storage systems play a critical role in supporting modern applications and meeting their performance, reliability, and scalability requirements. Current in-memory storage systems adopt three design decisions that limit their performance and efficiency. First, these systems rely on the write-ahead log to guarantee data consistency and tolerate failures. The write-ahead log enforces a strict sequential ordering on operations that is often unnecessary for many applications, introducing a performance bottleneck. Second, these systems are designed for traditional, server-centric hardware, overlooking potential design optimization of emerging hardware capabilities and rendering them incompatible with the recently proposed hardware-disaggregated architecture. Third, their disaster recovery mechanisms are designed under the assumption of complete time asynchrony across machines, resulting in either a large data loss window or a significant performance overhead. This thesis explores a fundamentally different design space for building high-performance, replicated in-memory storage systems.

First, to address the inefficiency of the write-ahead log, this thesis explores a novel system design that forgoes the write-ahead log and builds the Logless, Linearizable Key-Value storage system (LoLKV). By removing the log, LoLKV eliminates the serialization bottleneck and unnecessary memory copy operations, achieving a higher level of concurrency and improving resource utilization. LoLKV relies on one-sided RDMA to efficiently replicate data. Evaluation results demonstrate that LoLKV achieves 1.7–10× higher throughput and 20–92% lower tail latency compared to state-of-the-art RDMA-based systems.

Second, to address the performance challenges of current storage systems on the hardware-disaggregated architecture, I propose SplitKV, a low-latency linearizable key-value store designed for the hardware-disaggregated architecture. SplitKV leverages one-sided RDMA for communication with memory nodes, ensuring that memory nodes remain completely passive. SplitKV co-designs the replication protocol with the data structures of the system to minimize the number of RDMA operations required to process client requests. Evaluation results show that SplitKV achieves 2.6–21× higher throughput and 80–89% lower latency compared to Sift, the state-of-the-art disaggregated key-value store.

Finally, to address the shortcomings of current disaster recovery mechanisms, I leverage modern data center time synchronization hardware and protocols to build Slogger, a new disaster recovery system. Slogger achieves near-zero data loss and guarantees prefix linearizability at the backup site. Slogger uses continuous asynchronous replication to minimize the overhead on the system. Slogger employs a watermark service to guarantee the linearizability of the backup site while avoiding across-shard coordination. Evaluation experiments show that Slogger reduces the data loss window by 50% compared to the incremental snapshotting approach.