HIGH-PERFORMANCE NETWORKING IN DISTRIBUTED QUASI REAL-TIME SYSTEMS

This is a report of the work done during the Openlab Summer Student Programme 2022 within the ATLAS DAQ Dataflow group. The goal of the project was to characterize the baseline network performance of a 100GbE capable test bench, constituted of 8 servers interconnected by a switch. A second phase required to investigate the parameters of the network protocol in use, namely TCP/IP, and the network card capabilities, tuning them in order to improve the obtained results. The main utilized tools are the iperf command to run network tests, Python to execute multiple instance of iperf and collect data, and matplotlib library for plotting purposes. Benchmarks where run for single and many-to-one connections, to measure average throughput and find possible bottlenecks. Network patterns are governed by flow and congestion control of the TCP protocol which tries to ensure a fair bandwidth sharing to all connections, and, and as their number increases, so does packet loss. Setting a congestion control algorithm for data center environments (like dctcp), reducing TCP retransmission timeout and interrupt coalescing are some of the analyzed techniques which can help improving aggregated throughput and/or make single connection rates more stable