The invent of modern RDMA-enabled networks in data centers calls for a fundamental re-evaluation of how distributed data stores should be designed. Conventionally, many data structures, algorithms, and on the higher level, the architecture of these systems were primarily centered around the assumption that network communication is the dominant bottleneck and thus should be avoided as much as possible. However, in the light of recent advances in modern interconnects, this assumption is no longer valid, as RDMA functionality, together with high bandwidth and low latency of these networks enables direct remote access with significantly more efficiency and lower overhead compared to conventional networks. In this work, we first present a novel distributed OLTP system and show that it can truly scale even when all transactions span multiple machines. Motivated by this insight, we then argue that the conventional data partitioning techniques are no longer optimal in this new environment. To that end, we propose a new data partitioning scheme that optimizes for contention, which is the real bottleneck in the next generation of transactional databases. Finally, we tackle the problem of high availability in OLTP systems, and propose a new replication scheme that efficiently leverages RDMA to completely eliminate the processing redundancy in existing techniques.
Zamanian Dolati, Erfan,
"Distributed Transaction Processing on Modern RDMA-enabled Networks"
(2020).
Computer Science Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://doi.org/10.26300/8c6e-h719
