An improved deadline-aware scheme with low overhead in datacenter networks

Recently, datacenters have become increasingly important to provide a diverse range of online web services and cloud computing resources. As opposed to the pay-as-you-go model which cloud computing resources follow, user experience claims for specific service requirements and ultimately determines the revenue of online web services. The soft real-time nature of the large scale online web services with distributed workflow in today's datacenters, as one of the key challenges, leads to deadline being a critical requirement associated with the datacenter application traffic. By analyzing the existing deadline-aware rate decision protocols, we find that there is often a part of bandwidth consumed by some unproductive deadline constrained flows. What's more, D2TCP, as the most optimal protocol on meeting deadlines, incurs the most expensive overhead as well. In this paper, we propose an improved deadline-aware scheme with low overhead (DSLO) for datacenter networks to restrict the bandwidth overhead of the unproductive flows as well as the number of paralleled deadline constrained flows. We have evaluated DSLO under a typical fat-tree topology by ns-3 simulation tools. Our preliminary results show that DSLO can effectively reduce the overhead and deadline missing rate simultaneously.