SHF: EAGER: HI-HDFS - Holistic I/O optimizations for the Hadoop distributed filesystem

点击翻译按钮获取中文摘要

File systems and their outdated POSIX "byte stream" interface suffer from an impedance mismatch with the versatile I/O requirements of today's applications. Specifically, the I/O path from the application to the raw storage device is becoming longer and it involves the interplay of intricate software and hardware components. This produces complex aggregate I/O patterns that application developers (often subject matter experts with limited knowledge of how massive concurrency creates I/O bottlenecks) cannot optimize based on intuition alone. File systems that tout their high scalability, such as the Hadoop distributed file system, largely do so by limiting applications to sequential access patterns. The question of whether one can accelerate the I/O performance of the Hadoop distributed file system for analytical applications with complex data models that cannot readily serialize data contiguously for fast sequential access remains open. This project seeks to address this question and build HI-HDFS -- a framework that automatically collects and manages semantically richer I/O metadata to guide object placement in the Hadoop distributed file system. The HI-HDFS framework synthesizes the I/O activity across software components throughout the datacenter in a navigable graph structure to identify application-agnostic motifs in I/O activity. A novel I/O forecasting technique identifies and ameliorates bottlenecks at large scale by inspecting I/O activity from small-scale runs. Overall, the HI-HDFS framework challenges the I/O optimization mantra that manual data placement is the cornerstone of I/O performance and paves the way towards next-generation object-centric storage systems for high-performance computers. The efficacy of this automated approach will be examined on a complex data processing workload from the domain of emergency response which exhibits I/O patterns that are characteristic of modern analytical applications. The broader impacts of this work are expected to include open-source prototype implementations as well as pedagogical impact on a cloud computing course for both Computer Science and Data Analytics undergraduate majors at Ohio State.