On Scalability of Two NoSQL Data Stores for Processing Interactive Social Networking Actions

This paper quantifies the scalability of a document store named MongoDB and an extensible store named HBase for processing simple operations using a social networking benchmark named BG (www.bgbenchmark.org). We report on the vertical and the horizontal scalability of both data stores. We quantify speedup and scaleup characteristics of each data store's Social Action Rating (SoAR). SoAR is the highest observed throughput with a data store while satisfying a Service Level Agreement (SLA) such as 95% of requests observing a response time of 100 milliseconds or faster. While the speedup experiments maintain a fix sized social graph and vary the number of nodes, the scaleup experiments vary both the size of the social graph and the number of nodes proportionally. A system provides a super-linear SoAR speedup as a function of the number of nodes when it transitions from fully utilizing its disk bandwidth to fully utilizing either its CPU cores or network bandwidth. The speedup and scaleup of both data stores is limited with a resource such as the network interface card of one to three nodes becoming fully utilized and dictating the SoAR of a many (12) node deployment. One may configure MongoDB to use replicas of shards, enhancing its speedup characteristics and incurring a very small amount (<0.0003%) of unpredictable data with workloads consisting of writes.