CSR---PDOS: A Benchmarking Framework for High-Availability Distributed Storage Systems

I/O系统的可用性和健壮性对于生成和分析TB级数据的大规模应用程序至关重要。存储系统容易受到许多硬件故障（I/O和元数据服务器崩溃）的影响，并占所有系统故障的25%。实际上，随着高端计算系统规模的扩大，用于高端计算的高可用数据存储变得越来越重要。为了实现高可用性存储系统，一个具有挑战性的问题是描述除了这些系统的性能的可用性度量。调查人员采取有组织的方法来开发基准框架，以衡量存储性能，同时考虑各种故障条件下的可用性。研究内容包括四个方面：1）建立存储系统的故障/错误模型，设计故障注入方案; 2）建立不同故障条件下高可用分布式存储系统的基准测试框架; 3）实施可用性和性能评估工具集（APET）集成故障注入和压力测试库，捕获存储系统在各种故障下的块级原始性能; 4）使用APET对块级存储系统的基准测试框架进行验证。该研究对于理解HEC系统的高可用数据和I/O服务，建立故障条件下存储系统的通用基准测试框架，从而指导开发面向高可用性的分布式存储系统具有直接的贡献，这对许多应用至关重要。

The availability and robustness of the I/O system is crucial to large-scale applications that generate and analyze terabytes of data. Storage systems are vulnerable to numerous hardware failures (I/O and metadata server crashes) and contribute to as much as 25% of all system failures. Actually, highly available data storage for high end computing is becoming increasingly more critical as high-end computing systems scale up in size. To achieve high availability storage systems, a challenging issue is to characterize the availability metric in addition to performance of these systems.This research investigates high-availability data and I/O services and benchmarking. The investigators take an organized approach to developing a benchmarking framework to measure the storage performance in consideration of availability under various faulty conditions. The research involves four tasks: 1) develop faults/errors model and design fault injection schemes for storage systems; 2) develop an innovative benchmarking framework for high availability distributed storage systems under different faulty conditions; 3) implement an Availability and Performance Evaluation Toolset (APET) to integrate the fault injection and stress testing libraries and capture raw performance of storage systems at block level under various faults; 4) validate the benchmarking framework using APET for block-level storage systems. This research has direct contributions to understanding highly available data and I/O services for HEC systems, establishing a general benchmarking framework for characterizing storage systems under faulty conditions, and thus benefiting the society by guiding develop high-availability oriented distributed storage systems which are crucial to many applications.