Semester of Graduation
Electrical and Computer Engineering
First Major Professor
Master of Science (MS)
Modern storage systems utilize Non-Volatile Memories (NVMs) to reduce the performance and density gap between memory and storage. NVMs are a broad class of storage technologies, including flash-based SSDs, Phase Change Memory (PCM), Spin-Transfer-Torque Random Access-Memory (STTRAM). These devices offer low latency, fast I/Os, persistent writes, and large storage capacity compared to volatile DRAM. However, researchers are still working on the possibility of building systems that can leverage these NVMs to deliver low latency and high throughput to applications. Conventional systems were designed to persist data on hard drives, which has higher latency than NVM devices. Hence, in this work, we intend to explore opportunities to improve performance and reliability in the NVM based systems. One class of NVM devices that are placed on the memory bus is Persistent Memory (PM). Examples of PM technologies include 3D XPoint, NVDIMMs. Applications need to be modified to use the PM devices, which requires a lot of human effort and could lead to programming errors. Hence, reliability is also necessary to build systems to utilize the PM. Additionally, as persisted data is expected to be recoverable systems in case of a crash, PM applications are responsible for providing that reliability support at the application level instead of relying on the file system.
In this work, we evaluate the performance of popular key-value store RocksDB that is optimized for flash storage and also the reliability guarantees provided by recent works, which provides the testing framework for determining crash-consistency bugs in PM systems. Based on this analysis, we also present some opportunities to optimize performance and reliability in NVM systems.
Embargo Period (admin only)
Bansal, Prakhar, "Exploring Optimization Opportunities in Non-Volatile Memory Systems" (2020). Creative Components. 627.