Today’s cloud storage systems lack flexible cost-performance trade-offs. For example, (a) in database systems, there are only a limited number of cost-performance options and they are not seamless, (b) in cloud caching systems, there is no flexibility in performance isolation, and (c) in geo-replication systems, the cost-performance trade-off is not optimal to various application types. In this thesis, we develop novel mechanisms that offer greater flexibility for making finer, online cost-performance trade-offs for data storage systems using (a) data access statistics and (b) models that capture information regarding cost and user experience. We specifically look at ways of achieving better cost-latency trade-offs in the following problem domains: (Mutant) NoSQL database systems, (SpaceLease) cloud caching systems, and (Acorn) geo-replicated, multi-data center systems. With NoSQL database storage systems, we observe the inflexibility in the cost and performance trade-offs: the trade-offs have limited options and the transition between different cost-performance points are not automatic. We address the inflexibility by proposing Mutant, a NoSQL database storage layer that seamlessly trades off cost and performance. We implemented Mutant by modifying RocksDB, a popular NoSQL database, and evaluated with both synthetic and real-world workloads to demonstrate the seamless and automatic cost-performance trade-offs. With edge cloud caching systems, we observe the unpredictable performance in public cloud cache services: CPs (content providers) pay the same amount of price, but they get unstable cache hit rate over time. We address the performance unpredictability by proposing SpaceLease, a performance-isolated cache architecture that uses dedicated resource for caching data in the edge cloud platform. We implemented SpaceLease and showed up to 79% reduction in the performance variability with a minimal cost overhead. In addition to the stable performance, SpaceLease also (a) provides a control that trades off cost and hit rate, (b) maximizes the aggregate cache utility across data centers, and (c) adapts quickly to changing workload patterns. With geo-distributed multi-data center replication systems, we observe that (a) better replication decisions can be made by using the “right” object attribute for each application type, such as topics for public video sharing applications and users for social network applications, and (b) using the combinations of the attributes and extra random replicas makes better replications under a cost or latency constraint. In response, we developed Acorn, an attribute-based partial geo-replication system, and showed that Acorn delivers up to a 90% cost reduction or a 91% latency reduction.
PDF
download
878987797
028-85220240
