Date of this Version
The request latency is an important performance metric of a distributed database, such as the popular Apache Cassandra, because of its direct impact on the user experience. Specifically, the latency of a read or write request is defined as the total time interval from the instant when a user makes the request to the instant when the user receives the request, and it involves not only the actual read or write time at a specific database node, but also various types of latency introduced by the distributed mechanism of the database. Most of the current work focuses only on reducing the average request latency, but not on reducing the tail request latency that has a significant and severe impact on some of database users. In this thesis, we investigate the important factors on the tail request latency of Apache Cassandra, then propose two novel methods to greatly reduce the tail request latency. First, we find that the background activities may considerably increase the local latency of a replica and then the overall request latency of the whole database, and thus we propose a novel method to select the optimal replica by considering the impact of background activities. Second, we find that the asynchronous read and write architecture handles the local and remote requests in the same way, which is simple to implement but at a cost of possibly longer latency, and thus we propose a synchronous method to handle local and remote request differently to greatly reduce the latency. Finally, our experiments on Amazon EC2 public cloud platform demonstrate that our proposed methods can greatly reduce the tail latency of read and write requests of Apache Cassandra.
Adviser: Dr. Lisong Xu