HMM Optimized Modeling of SSD Storage for I/O MapReduce Workloads

Flash-based SSD draws a considerable interest in big data platforms due to its performance and reliability. However, it still has limited usage as a result of its high cost and limited capacity. Control SSD provisioning on big data platforms reduce storage cost and guarantees performance. The workload is an essential SSD provisioning sources, thus analyzing the characteristics of the workloads would help optimize SSD management design. There is a significant correlation between the workload's IO patterns and the SSD cost and performance. Big data platforms with multi-stage architecture bring challenges into modeling IO patterns where each stage has it is unique IO patterns. Also, big data platforms run on a distributed environment where the workloads are interacting with local and remote storage during the execution. The designed HMM-based IO patterns model considers IO patterns for MapReduce workloads at different stages and different SSD locations. In this paper, we proposed a platform-level SSD, cost-efficiency controller. The controller is responsible for maximizing the SSD lifespan on the Hadoop platform through two phases. First, modeling MapReduce workload's IO patterns by employing the Hidden Markov Model (HMM). Then, defining platform-level SSD allocation policies. The designed allocation policies reduce SSD utilization and improve SSD lifespan on Hadoop by up to %40 compared to static allocation policies.