Parallel

Computing

and I/O

• Author: Timm Leon Erxleben
• Type: Bachelor's Thesis
• Date: 2022-05-13
• Reviewers: Jun.-Prof. Dr. Michael Kuhn, Dr. Martin Köhler
• Download: PDF

Abstract

Ever-increasing demands on capacity and data throughput of storage clusters lead to growing systems and growing energy consumption. Currently, storage clusters are built from regular x86-based servers that are not energy-proportional and have a high idle power consumption. Most distributed file systems cannot fully utilize servers leading to low energy efficiency. ARM-based processors were designed for energy efficiency and low power consumption. Therefore, an ARM-based single-board computer storage cluster was built using the Odroid HC4. This cluster is evaluated using different distributed file systems, workloads, metrics, and a comparable x86-based setup as reference. Results show that ARM-based single-board computer storage clusters are an energy-efficient alternative to traditional x86-based clusters while providing comparable performance. However, based on the experiences with the Odroid HC4, future single-board computers for storage clusters should have a higher network throughput and should provide access to more storage devices.