Reducing Bandwidth Waste in Reliable Multistream Storage
Andromachi Hatzieleftheriou, Stergios V. Anastasiadis · 2011
Synchronous small writes play a critical role in the reliability and availability of current systems because they are used to safely log recent state modifications and allow fast recovery after failures at the application and system level. In highly demanding environments, it is typical to dedicate separate devices for the logging activity alone. Thus, systems provide adequate performance during normal operation and extra redundancy for state reconstruction after a failure. However, storage stacks usually enforce page-sized granularity in their data transfers from memory to disk. Under various conditions, we experimentally show that subpage writes may lead to storage bandwidth waste and high disk latencies. To address the issue in a journaled file system, we propose wasteless journaling as a mount mode that coalesces synchronous concurrent small writes of data into full page-sized blocks before transferring them to the journal. Additionally, we propose the selective journaling mode that automatically applies wasteless journaling on data writes whose size lies below a fixed preconfigured threshold. Depending on the request size, written data is either first coalesced to the journal of the file system, or directly transferred to its final location on disk. We developed a prototype implementation of our design based on a widely-used file system. Then, we apply microbenchmarks and application-level workloads to standalone servers and a multi-tier networked configuration to show substantial improvements in handling small write traffic over write latency, transaction throughput, recovery time and storage bandwidth requirements.