A Predictive ECMP Routing Protocol for Fat-Tree Enabled Data Centre Networks

Abstract

Due to the exponential growth of cloud computing, data centres have become the pivot for supporting the core infrastructure that propels the cloud computing evolution. Data centres are repositories that house different networking devices that are connected together using communication links to collect, store, process and disseminate data. Data centres prioritise high data availability amongst others. However, data availability is challenged by the unpredictable nature of the network environment, which presents enormous challenges in designing routing protocols that are agile enough to adjust to sudden changes in the network's available capacity. To provide seamless services to users, most modern data centres use Fat-Tree as the de-facto topology due to its multipath diversity, and the Equal-Cost Multi-Path protocol (ECMP) as the primary routing protocol to route data towards alternative paths of equal cost when the primary path is over-utilised. However, the weighted algorithm used to achieve this is inefficient, as its assigns traffic to links based on their link capacities without taking into consideration the capacity already in use on that link. In this paper, we propose the Predictive Equal-Cost Multi-Path protocol that enhances ECMP's weighted load-balancing algorithm by making forwarding decisions based on predicted congestion outlooks. The proposed protocol uses packets dropped to compute the bandwidth utilisation of links and uses the computed figures to identify the least congested links, which is then used to build forwarding tables. The protocol was tested in a Fat-Tree enabled data centre where it proved to perform better when compared to the ECMP weighted algorithm.