Abstract
InfiniBand is one of the most popular interconnection network standards, in the area of High-Performance Computing (HPC). The InfiniBand architecture supports deterministic routing. This prevents packets from using alternative paths when the requested output port is busy, which consequently degrades the performance of the system. Thus, adaptive routing algorithms, which dynamically select the route of a packet, need to be considered. An output selection function is essential for an adaptive routing mechanism, as it determines the output channel when some valid channels are free. Although, many adaptive routing algorithms have been studied, comparatively little work has been done on determining the selection function in infiniband. In this thesis, three output selection functions are proposed for use in infiniband, and their impact on the network performance have been studied. The first mechanism is based on Round-Robin (RR) policy, the RR uses a round robin algorithm to choose a different physical link each time without considering the link status. The second and third selection mechanisms use a local information available in the switch to select the output link; Maximum Credits Available (MC) selects the port with the most available credits in the buffer, while Less Busy Channel-Least Recently Used (LBC-LRU) has the following strategies for distributing the traffic: 1) each switch locally grasps the congestion information on the output ports; 2) choose from a set of less busy channels according to LRU policy. The proposed selection functions can be used on any type of network topology and adaptive routing algorithm in infiniband. However, we have decided to simulate our proposals on fat-tree topologies because of fat-tree popularity among switch manufacturers, and its guarantee for deadlock- freedom. The simulation results show that the choice of selection function has a significant impact on the average message latency, network throughput and variation in packets delay. All three functions show an improvement in the network performance. The LBC-LRU selection function provides the best performance, especially when the network size becomes larger.