Taming Unaligned Writes in Solid State Disk
Abstract
NAND-flash based solid-state disks (SSDs) are replacing the hard-disk drives (HDDs) in various storage systems from high-end servers in data centers to mobile computers on the edges of cloud computing. Due to the architectural nature of SSDs, performance-sensitive applications may generate a large amount of unaligned writes on SSDs, which can cause many issues (e.g., chip congestion, sub-request blocking, chip load imbalance, and write space amplification). We present the design and implementation of Partial Congestion aware LRU Algorithm (PC-LRU) for Write Buffer Management that comprehensively and significantly alleviates the side-effects of unaligned writes in solid-state disks. This technique prioritizes the eviction of fully-filled pages over partially-filled pages in write buffers to circumvent write amplification. We demonstrate the advantages of using trace-driven simulations in SSDSim. For 3 real I/O workloads, PC-LRU obtains average latency improvement by up to 9.5%, 19% for scaled timestamps and aging condition respectively.
About the Presenters
Abu Zafar Md Nuruzzaman Abir
Abu Zafar Md Nuruzzaman Abir is a graduate Researcher working in Parallel and Distributed Storage Systems Lab who is pursuing M.S. in Computer Science. His focus is on the solid-state storage system and he gained much experience working in the lab focusing on exploring different storage systems. He has a working knowledge of developing projects based on Cloud storage systems. Furthermore, he has also developed a project focusing on reducing the number of writes in the buffer as part of the File and Storage System course.