Performance measurement of SimpleDB APIs for different data consistency models

Cloud platform providers usually offer several APIs (Application Program Interface) to help facilitate programmers to utilize cloud resources effectively by hiding complex cloud structures and mechanisms. However, there are some aspects of distributed computing that cannot be hidden. Depending on data consistency models, two or more clients may not see the same current state of the data. Developers should understand the performance of available APIs and consistency models they provide. This paper explores the performance of a suite of APIs that can be used to implement two different data consistency models in SimpleDB, a distributed non-relational database-as-a-service provided by Amazon. Based on our experiment, read requests in two consistency models offered by Amazon SimpleDB performed almost identically, with median latency of 16 ms. Write performance was about 3 times slower, with 56 ms median latency. In addition, there were greater performance variations for writes than reads. Lastly, a strong consistency model worked as advertised, returning latest value with every read. On the other hand, the correctness of an eventual consistency read depended primarily on the elapsed time since last write operation.