Volume 18, No. 2, 2021
HCCLBA: Hop-By-Hop Consumption Conscious Load Balancing Architecture Using Programmable Data Planes
V. Rajkumar , Dr. V. Maniraj
Abstract
Datacenter networks typically make use of multi-rooted topologies in order to deliver enormous bisection bandwidth (such as leaf spines and fat trees). In order to make the most of the available bisection band width, it is necessary to have a load-balancing mechanism for the data plane. This is due to the significant degree of multipathing that is present in these topologies. The most typical approach of load balancing is called equivalent-cost multi-path routing, or ECMP for short. This method moves traffic down a number of different routes simultaneously. Congestion Aware Load Balancing Techniques, such as CONGA, have been created as a solution to the restrictions that ECMP presents. Using these approaches does come with a few of limitations. To begin, the amount of congestiontracking state that can be retained at the edge switches is restricted due to the memory limitations of the switches, which makes it hard to scale complicated topologies using these switches. The second drawback is that because they are incorporated into specialized hardware, they cannot be modified locally on the premises where they are being used. A load-balancing approach for the data plane in this research, HCCLBA is offered as a solution to overcome both of these concerns. As a first stage, rather of watching congestion on all possible pathways, each HCCLBA switch only tracks congestion on the most efficient path to the destination, which is typically through an adjacent switch. It is also feasible to run HCCLBA on these chipsets without the requirement for new hardware by designing it to work on programmable switches and programming it in P4. This would make it viable to run HCCLBA on these chipsets. The simulation shows that HCCLBA has a faster average flow completion time than a scalable extension to CONGA (1.6 times faster at 50 percent load and 3 times faster at 90 percent load).
Pages: 1985-1995
Keywords: HCCLBA, Load Balancing, Software Defined Networking.