Efficient qos support for high-performance interconnects
José Luis Rodríguez García, Francisco J. Alfaro Cortés, Alejandro Martínez Vicente · 2008
Interconnection networks are a key component of a variety of systems. Today, low-latency and contention-free interconnection networks are demanded for the execution of many applications in systems like supercomputers, clusters of PCs, and others. There are many issues involved in successful high-performance interconnects. Among them, quality of service (QoS) is responsible of guaranteeing that a certain performance is achieved. Depending on the specific applications, guarantees on latency, throughput, and other indices may be critical. Traditional solutions to provide QoS in high-performance interconnects usually rely on complex architectures. For instance, virtual channels are a mechanism that is often proposed to isolate several traffic classes. However, it is unusual to see in a final implementations as many virtual channels as in proposals. The main objective of this thesis is to investigate if we can offer efficient mechanisms to provide QoS. Our purpose is to achieve a full QoS support with the minimum of resources. We do that by identifying redundancies in current proposals for QoS support and eliminating them. This thesis consists of three parts. In the first one we take as a starting point the traditional proposals of QoS at the traffic class level. We propose a more efficient framework for QoS provision using just two virtual channels at the switches. This is possible because we take advantage of scheduling performed at end-nodes. In the next part, we propose how to adapt deadline-based QoS algorithms to high-performance interconnects. Using an approach similar to that of the previous part, we obtain remarkable results with simple switch architectures. Finally, we investigate in the interaction of QoS mechanisms and congestion management techniques. We propose an integrated solution that benefits from the synergy between both aspects of high-performance networking.