Computer Science Analysis and optimization of a high-speed InfiniBand messenger

With the advent of modern many-core processors and the wide spread use of GPUs as co-processors, raw computing power has become almost abundant. However, as the number of physically separated processing units in a system increases, so does the requirement for efficient interconnects between those components.

In the field of supercomputers, InfiniBand has become the most used network interconnect. At IPE, we have developed a communication library for InfiniBand, called KIRO, which is used to transfer data between the nodes of a GPU computing cluster. It leverages InfiniBands RDMA (Remote Direct Memory Access)capabilities to reach bandwidths of multiple gigabytes per second and latencies in the order of a few microseconds. Due to these very low latencies, even small delays in the KIRO software Event handling have an immediate impact on InfiniBand's performance. Optimizing KIRO would help to increase the efficiency of the network used in the computing cluster and thus allow for lower response times and higher scalability.

For this proposal, you are supposed to:
• Compare the implementation of KIRO's event handling and dispatching mechanism with the current state-of-the-art
• Propose a strategy to improve KIRO's request-response turnaround times
• Verify the feasibility of the proposed improvements for Linux and Windows alike

This project will be embedded into the international UFO project for ultra-fast tomography with online reconstruction and fast feedback control loops.

Weitere Informationen

Unternehmen

Helmholtz Gemeinschaft

Bereich/Abteilung

Institute for Data Processing and Electronics (IPE)

Abschlussart

Bachelorarbeit

Branche

Informationstechnologie und -dienste

Anforderungen

Solid knowledge of the C programming language as well as Linux and Windows software development. Prior knowledge of networking and Remote Direct Memory Access (RDMA), is a plus.

Experience gained:
InfiniBand interconnects, RDMA data transfer mechanisms and high-performance networking.

Zusatzinformationen