Dynamic Distributed Flow Scheduling with Load Balancing for Data Center Networks

Abstract

Current Flow Scheduling techniques in Data Center Networks(DCN) results in overloaded or underutilized links. Static flow scheduling techniques such as ECMP and VLB use hashing techniques for scheduling the flows. In case of hash collision a path gets selected number of times resulting overloading of that path and underutilization of other paths. Dynamic flow scheduling techniques like global first fit employ centralized scheduler and always selects first fittest candidate path for scheduling. Thus in addition to single-point-of-failure the overall link uti- lization also remains a problem as the flows are not scheduled on the best available candidate path. This paper presents firstly a Dynamic Distributed Flow Scheduling(DDFS) mechanism that will lead to fair link utilization in globally used fat-tree topology of DCN. Secondly, it presents a mechanism to restrict the flow scheduling de- cisions to the lower layers thus avoiding saturation of core switches. The entire DCN is simulated using Colored Petri Nets (CPN). The load measured at the aggregate switches for various flow patterns in DCN reveals that the load factors at the aggregate switches vary by at most 0.11 which signifies the fair utilization of links.