Prioritizing deadline-constrained data flows in cloud datacenter networks

Abstract

Real-time interactive application workloads (e.g. web search, social networking, etc.) are composed of a remarkably large number of mini request partitions that require stringent delayminimal aggregation of responses and their delivery back to the user. This paper presents a novel Time-Deadline-AwarepFabric (TDA-pFabric) being a data transport design whose objectives are two-fold: i) reducing the mean Flow Completion Time (FCT), the deadline mismatch and blocking probabilities as well as ii) improving the mean application throughput. A stochastic queueing model is proposed herein to capture TDA-pFabric's dynamics and evaluate its performance. The model is validated through extensive simulations whose results also show that TDA-pFabric exhibits significant improvements over existing designs in terms of the above-listed quality-of-service metrics.