Analyzing the Performance of the S3 Object Storage API for HPC Workloads

The line between HPC and Cloud is getting blurry: Performance is still the main driver in HPC, while cloud storage systems are assumed to offer low latency, high throughput, high availability, and scalability. The Simple Storage Service S3 has emerged as the de facto storage API for object storage in the Cloud. This paper seeks to check if the S3 API is already a viable alternative for HPC access patterns in terms of performance or if further performance advancements are necessary. For this purpose: (a) We extend two common HPC I/O benchmarks—the IO500 and MD-Workbench—to quantify the performance of the S3 API. We perform the analysis on the Mistral supercomputer by launching the enhanced benchmarks against different S3 implementations: on-premises (Swift, MinIO) and in the Cloud (Google, IBM…). We find that these implementations do not yet meet the demanding performance and scalability expectations of HPC workloads. (b) We aim to identify the cause for the performance loss by systematically replacing parts of a popular S3 client library with lightweight replacements of lower stack components. The created S3Embedded library is highly scalable and leverages the shared cluster file systems of HPC infrastructure to accommodate arbitrary S3 client applications. Another introduced library, S3remote, uses TCP/IP for communication instead of HTTP; it provides a single local S3 gateway on each node. By broadening the scope of the IO500, this research enables the community to track the performance growth of S3 and encourage sharing best practices for performance optimization. The analysis also proves that there can be a performance convergence—at the storage level—between Cloud and HPC over time by using a high-performance S3 library like S3Embedded.