An I/O Performance Evaluation of Varying CephFS Striping Patterns

• Published in: 2023 IEEE International Conference on Cluster Computing Workshops (CLUSTER Workshops) pp. 25 - 31
• Main Authors: Biswas, Debasmita, Neuwirth, Sarah, Paul, Arnab K., Butt, Ali R.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 31.10.2023
• Subjects: Big Data; Ceph Stripe Configuration; CephFS; Cluster computing; Computer architecture; Conferences; File Striping; File systems; High Performance Computing; I/O Performance; Layout; Performance evaluation
• DOI: 10.1109/CLUSTERWorkshops61457.2023.00014

With the enormous rise in data creation and consumption, there is a persistent need to upgrade existing technologies that can keep up with the speed of big data generation, and facilitate faster I/O, a bottleneck in this area. Object Storage architectures thrive on the principle of storing data as objects attached with globally unique identifiers in a flat storage pool. This does away with the traditional tiered designs of Parallel and Block Storage Systems and facilitate faster read/write operations. The Ceph File System is built on top of the Ceph Object Store. CephFS utilizes the Raid 0 striping strategy to stripe incoming data over object(s) which are then mapped to the OSD(s). Currently, this striping strategy is determined at the time of cluster deployment and remains unchanged. By default, an object size and stripe width of 4MB with a Stripe Count of 1 is used for the CephFS. In this work we explore the Ceph Stripe configuration by varying the object size, stripe count and stripe unit, to gain initial observations of how these changes affect read and write I/O of HPC applications.