Radio+Tuner: A Tunable Distributed Object Store

The primary reason for diversity of distributed storage systems is that no single design for a distributed storage system is cost-optimal for meeting the latency and throughput SLOs in all possible workloads. In other words, a distributed object store that satisfies the SLOs for a particular workload at minimum cost (in comparison to other object stores) will incur higher costs than alternate storage systems on other workloads. Therefore, an application administrator today is faced with the onerous task of choosing from the large variety of available object stores; an incorrect choice may significantly inflate costs necessary to meet performance SLOs. Furthermore, as new workloads emerge, none of the existing systems may be a good match, thus necessitating the development of new object stores. This is because existing systems are inflexible, having assumptions of their target workload are baked into their implementation, and making it unclear how any particular system should be modified in order for it to be well-suited for a workload other than that for which it was designed. While any existing system can likely meet an application’s performance goals with infinite resources, we believe this is the wrong approach.