Physics Data Forge: Unveiling the Power of I/O Systems in CERNs Test Infrastructure

Abstract

Efficient remote file access is essential for High Energy Physics (HEP) experiments, particularly with the anticipated tenfold increase in data volume for Run 4 of the LHC. XRootD and EOS are critical components of the HEP ecosystem, enabling remote data access via XRootD and HTTP protocols. In this work, we evaluate the performance of various client implementations over these protocols using an XRootD server with a high-performance tmpfs mount. In order to ensure an infrastructure-independent assessment, benchmarking is conducted in a controlled 100GbE networking environment, eliminating external bottlenecks and isolating software performance. Network integrity was first validated using iperf3 before executing the benchmarks. Our study identifies a performance bottleneck in multi-stream data transfers at high rates within the XRootD client. A fix has been developed and will be included in XRootD 5.8.0, significantly improving high-throughput data access.