Abstract
The processing of large data volumes in forensic applications places high demands on the network performance of HPC clusters, especially regarding bandwidth and latency. In addition to upholding forensic principles, such as data integrity, rapid response times are essential for data-intensive analyses like topic modeling, file carving, and decryption. This thesis investigates the integration of a high-speed interconnect technology into an HPC cluster to enhance forensic applications and addresses how network performance impacts the efficiency of cluster operations.Following a detailed requirements analysis, InfiniBand HDR and 200G Ethernet with RDMA over Converged Ethernet were identified as potential options, assessed according to strategic, technical, and organizational criteria. InfiniBand HDR was ultimately selected for its high bandwidth, low latency, and native RDMA support, and was integrated into the cluster. The integration process proceeded in multiple phases: initial hardware testing, installation and configuration, and software adaptation. To assess performance, RDMA benchmarks were conducted on read, write, and send operations across Intel and AMD nodes running Linux and Windows. The results show significant performance differences between systems, with nodes featuring PCIe 4.0 and Linux achieving superior bandwidth and message rates compared to PCIe 3.0 and Windows nodes. The new interconnect provides significant advantages for data-intensive workloads with up to 200 Gbit/s bandwidth and 600 ns latency, outperforming the interconnect previously used in the HPC cluster at the Secure Information Systems Department, University of Applied Sciences Upper Austria.
The findings of this thesis provide a foundation for future interconnect integration and benchmarking projects in the HPC field. By implementing RDMA in applications and increasingly adopting PCIe 4.0 interfaces, performance for forensic applications can be further enhanced, maximizing the potential of HPC environments.
Date of Award | 2024 |
---|---|
Original language | German (Austria) |
Supervisor | Thomas Grurl (Supervisor) |