Leveraging Heterogeneous Storage for Faster and Cheaper Computing Infrastructure

In recent years, several promising storage technologies have been developed with the goal of accelerating the performance and reducing the cost of computing infrastructure. On one end of the spectrum, high performance non-volatile memory (NVM) technologies provide latencies, which approach the speed of DRAM. On the other end of the spectrum, cheap and scalable flash storage, Quad Level Cells, or QLC, enables applications to store vast amounts of data on flash at a low cost per bit. QLC is at the price range where it becomes cost competitive with hard disk drives (HDD), which are much slower. However, in order to make it more cost effective, QLC trades off its endurance which is much lower than traditional flash. Therefore, there has been a large amount of recent work in building new databases, storage systems and file systems that take advantage of these technologies. However, these systems are built for a single class of storage technology. Therefore, they need to compromise: they are either very fast but expensive NVMbased systems, or very cheap but slow and unreliable QLC-based systems. In order to get the best of both worlds, systems will have to combine multiple heterogeneous storage technologies in tandem: ``hot'' and frequently updated data will be stored in faster, more expensive storage mediums, while ``cold'' and more static data will be stored in cheaper but slower storage technologies. We envision that in the future, critical computing infrastructure will transparently and automatically take advantage of a gamut of storage and memory technologies, which will be managed and scaled independently. The goal of my research proposal is to realize this vision, by designing systems that can exploit this heterogeneity, and achieve emph{both} the latency and throughput of one end of the storage spectrum, while retaining the cost of the other end of the spectrum.