Quantification and analysis of performance fluctuation in distributed file system

Performance is a key requirement in quality of service for cloud storage system users. Cloud storage system is complex, so the performance results may be different even under exact same experimental conditions. Many benchmarks are developed to measure performance, but there are only several research focus on performance fluctuations. In this paper, we present a generic method and two metrics to measure the performance fluctuations of a distributed file system. It analyzes the IO path and finds related factors such as storage media, network, configurations and software components. We tested the performance fluctuations of Ceph under different storage media HDD, SSD, and different workloads Webserver, Fileserver, and Mailserver. We adjust the software configuration and software components to reduce the performance fluctuations. By using Filestore, it reduces 85.4% performance fluctuations than Filestore. By using EXT4 instead of XFS, it reduces performance fluctuations by 33.2%. But reducing performance fluctuations of lower components does not always reduce the performance fluctuation of distributed file system. We reduced the performance fluctuation of the local file system by 80.7%, but the performance fluctuation of DFS by 38.0%.

Automated summary

This paper presents a generic method and two metrics to measure the performance fluctuations of a distributed file system. By analyzing the IO path and identifying factors such as storage media, network, configurations, and software components, as well as making adjustments to software configurations and components, performance fluctuations can be reduced.