Smart Network Interface Controller Systems Software for High Performance Networks

点击翻译按钮获取中文摘要

High Performance Networks (HPNs) are a critical aspect to large-scale computing systems. Large-scale computing systems make modern life possible, with large-scale Cloud computers facilitating social media interactions, entertainment, online meetings and making our smartphones smarter and more versatile. They also contribute to High Performance Computing (HPC) that is utilized for Artificial Intelligence (AI), scientific simulations for understanding our universe and even simulations of COVID-19 transmission. Large-scale systems can be more productive with better resource management capabilities. Unfortunately, this is difficult to accomplish without consuming computing resources that the systems are using to accomplish their tasks (CPU time, memory, etc.). Furthermore, using computing resources for better computing system management disturbs the applications running on the large systems, potentially even slowing them down. However, recent advances in HPN Smart Network Interface Controllers (SmartNICs) can help by offloading the resource management burden from the high-performance compute resources that applications need to a new compute component in the system, the SmartNIC. Modern HPN SmartNICs are a result of ever increasing network speeds, which required moving HPN component design to cutting-edge silicon production techniques used for CPUs with large silicon die sizes. Electrical engineering concerns meant that HPNs now had interface controllers that had much more space for transistors than was needed to perform their networking functions. SmartNICs filled this void by adding compute capabilities into the "blank silicon" that was left available. However, taking advantage of this new compute capability is difficult. SmartNICs are challenging to program and have the inherent disadvantage of being separated from the main compute CPU environment. This requires novel software engineering to take advantage of SmartNICs and completely rethinking past systems software approaches. We will leverage SmartNICs to develop novel resource management systems software, avoiding the overhead issues that such approaches suffered from when they needed to share resources with the main compute tasks. We will also develop novel system software to take advantage of the placement of compute resources in the network, creating new distributed software that can co-ordinate communication with other compute nodes in the system without perturbing applications. We will develop revolutionary new failure recovery, data manipulation and staging techniques for SmartNICs, allowing existing HPC applications to seamlessly take advantage of SmartNIC system software performance benefits. All of these techniques can be used to significantly increase Canadian public HPC extreme-scale systems performance, increasing Canada's industrial HPC engineering simulation competitiveness while training 3 PhD, 2 MSc and 3 undergraduate students in a very high demand sector.