File system performance is critical for overall performance of Big Data workloads. Typically, Big Data file systems consist of dual layers; a local node-level file system and a global file system. This paper presents the design and implementation of MorphStore, a local file system design that significantly improves performance when accessing large files by using two key innovations. First, MorphStore uses a load-adaptive I/O access scheduling technique that dynamically achieves the benefits of striping at low load and the throughput benefits of replication at high loads. Second, MorphStore uses a utility-driven replication to maximize the utility of replication capacity by allocating replication capacity to popular read-mostly files. Experiments reveal that MorphStore achieves 8% to 12% higher throughput while using significantly less replication for workloads that access large files. If we consider the performance-capacity tradeoff of file systems built on static techniques such as JBOD, RAID-0 and RAID-1 MorphStore extends the Pareto frontier to achieve better performance at the same replication capacity.

Emerging solid-state storage media can significantly improve storage performance and energy. However, the high cost-per-byte of solid-state media has hindered wide-spread adoption in servers. This paper proposes a new, cost-effective architecture - SieveStore - which enables the use of solid-state media to significantly filter access to storage ensembles. Our paper makes three key contributions. First, we make a case for highly-selective, storage-ensemble-level disk-block caching based on the highly-skewed block popularity distribution and based on the dynamic nature of the popular block set. Second, we identify the problem of allocation-writes and show that selective cache allocation to reduce allocation-writes - sieving - is fundamental to enable efficient ensemble-level disk-caching. Third, we propose two practical variants of SieveStore. Based on week-long block access traces from a storage ensemble of 13 servers, we find that the two components (sieving and ensemble-level caching) each contribute to SieveStore's cost-effectiveness. Compared to unsieved, ensemble-level disk-caches, SieveStore achieves significantly higher hit ratios (35%-50% more, on average) while using only 1/7th the number of SSD drives. Further, ensemble-level caching is strictly better in cost-performance compared to per-server caching.