Adaptive Routing Strategies for Modern High Performance Networks

Abstract

Abstract

Improving Performance of Deterministic Single-Path Routing on 2-Level Generalized Fat-Trees

“…Not all nodes need to participate in a permutation communication: a node that is not involved can be considered as communicating with itself, which is not represented in by Definition 1. A bisect communication pattern is a permutation where half of the nodes are sending to the other half of the nodes; and each node is either a sender or a receiver [2], [4]. Note that we assume is an even number.…”

“…The metrics used include average bisect bandwidth ( ), average full permutation bandwidth ( ), average dissemination bandwidth ( ). Since the numbers of bisect patterns, full permutations, and dissemination patterns are very large, we resort to a statistical method similar to one used in [2], [4] to compute average bandwidth for each patterns. We will describe how we compute .…”

“…Definition 4 captures the situation when each logical process can be mapped to any physical node. These communication patterns are used to measure the application level performance for interconnection networks [2], [4], [10].…”

“…In this paper, we develop a novel single-path routing scheme for 2-level generalized fat-trees that is optimal in terms of worst-case permutation performance. By improving worstcase permutation performance, our algorithm also achieves much higher average-case performance on all of the three types of fat-trees (full bisection bandwidth, slimmed, and fatted) for common communication patterns that are used to evaluate application level performance of interconnection networks, including the bisect patterns [2], the full permutation patterns, and the dissemination (Bruck) patterns [1], [2]. Our evaluation results show that the new routing scheme improves averagecase performance of bisection patterns by up to 17.6%, full permutation patterns by up to 57.7%, and dissemination patterns by up to 57.7%.…”

Channel reservation protocol for over-subscribed channels and destinations

“…Lossless flow control benefits performance, quality of service (QoS), and predictability over lossy (dropping) flow control networks, such as with Ethernet and TCP [24]. Lossy networks also typically have no router-to-router feedback mechanisms, which limits their adaptive routing capability [20,24,4]. Despite these benefits, lossless flow control suffers from congestion which can form multi-hop paths of blocked packets affecting benign flows, known as tree saturation [30,40].…”