Towards Application-aware In-network Bandwidth Management in Data Centers

A wide variety of applications nowadays are running on data centers, making the scarce in-network bandwidth of data centers become the performance bottleneck. To guarantee the network performance, an efficient in-network bandwidth management can be leveraged, however little work has been done so far. The crux is that existing work mainly relies on a static or dynamic bandwidth allocation strategy, which lacks efficient schemes to allow elastic and flexible use of the in-network bandwidth among different applications. In this paper, we study the in-network bandwidth management problem in data centers with a large number of diverse applications. To this end, we first formulate a max-min bandwidth constraint model to reserve an amount of bandwidth for each type of application. With this model, bandwidth can be sharing across different application types in an elastic and flexible way. We further present a Borrow-Return-Preempt bandwidth management method (BRP) to practically allocate bandwidth to individual requests of each type of applications. Extensive simulation results have shown that our proposed BRP is capable of improving the in-network bandwidth utilization as well as providing desirable network performance for applications, compared to the conventional fair allocation method.