Abstract
Current static flow scheduling techniques such as Equal Cost Multi Path (ECMP) and Valiant Load Balancing (VLB) use randomization and hash based techniques to schedule the flow in Data Center Networks (DCN). In such flow scheduling techniques, the probability of allotment of the same path for different flows at a time is high. This causes collision of flows and at times can lead to unfair link utilization. On other hand, Dynamic flow scheduling techniques such as Global first fit use a centralized mechanism to schedule flows. Global first fit can introduce the problem of unfair link utilization and also vulnerable to a single point of failure. Effective flow scheduling is the key problem in fat-tree topology based DCNs mainly due to the bandwidth over-subscription ratio and bursty nature of traffic. In this paper, we propose a dynamic distributed flow scheduling mechanism for effective link utilization and load balancing. Moreover our proposed mechanism prevents the core switch saturation scenario by taking flow scheduling decisions at lower layers. Load measurement performed on aggregate switches for various data center traffic flow patterns reveals that the load factors across all the aggregate switches vary by at most 0.11 when dynamic distributed flow scheduling is applied without flowlet splitting. Further with flowlet splitting our dynamic distributed flow scheduling mechanism improved the load factor variation to 0.08. The load factor values signify there is considerable improvement in link utilization and the switches are fairly load balanced. The experiment scenario is simulated using Colored Petri Nets.