Contention-free complete exchange algorithm on clusters

Abstract

To construct a large commodity clustec a hierarchical network is generally adopted for connecting the host muchines, where a Gigabit backbone switch connects a few commodity switches with uplinks to achieve scaled bisectional bandwidth. This type of interconnection usually results in link contention and has congestion developed at the uplink ports. Moreover, the non-detenninistic delays on scheduling communication events in clusters accelerate the building up of congestion amongst these uplink ports, which lead to severe packets drop and hinder the overall performance. In this paper, we focus on the practical design of high-speed complete exchange algorithm on a commodity cluster interconnected by a hierarchical Ethemet-based network. By exploiting some architectural characteristics of the interconnection in optimizing the performunce of a complete exchange algorithm, we introduce a congestion control mechanismglobal windowing that monitors and regulates the trafic load, together with a permutation scheme reorder scheme that effectively alleviates the congestion problem. We evaluate our algorithm and compare its performance with other algorithms in a PC cluster connected by various types of switches, including Gigabit Ethernet, input-buffered and shared-memory Fast Ethernet switches. 57 0-7695-0896-O/OO $10.00 0 2000 IEEE

Network performance variability in NOW clusters

“…Both techniques however require additional and modified hardware. Schemes such as overlays [21] and exchanges [26] have also been proposed. These however have the same effect of reducing network capacity or adding to the idle time of the process and application, potentially contributing to run-time variability of applications that are sensitive to these perturbations.…”

Exploring the relationship between parallel application run-time variability and network performance in clusters

“…Other approaches that attempt to address high performance network congestion include using what are commonly called overlays, in this case called multirail networks [30] and exchange schemes [34] that take into account the network topology to compute a communication schedule that is a contention free complete exchange of information in the cluster. Yet another mechanism proposed in [3] assumes the existence of virtual channels and estimates network traffic locally based on a percentage of free virtual channels available.…”