Traffic engineering in data center networks : prediction and scheduling via randomized rounding for elephant flows

Abstract

Applications and services hosted in large Data Centers account for most of the increase in Internet traffic. Data Center Networks (DCNs) are often designed with a fat-tree topology, allowing multiple paths between any two servers. The most widely adopted solution for flow routing in DCNs is equal-cost multipath (ECMP), which can cause link performance degra dation due to the possible occurrence of hash collisions in the presence of flows with many gigabytes of data, called elephants. Such collisions can result in packet discard, which generates packet retransmission, causes additional latency, and further degrades link performance. This thesis proposes a hybrid prediction model by combining aspects of the FARIMA and the Recur rent Neural Network (FARIMA-RNN) models to predict elephant flows on a short-term basis. Besides, we implement an SDN solution based on a randomized rounding heuristic, named RDRH, to schedule elephant flows in DCNs. We employ a linear programming formulation that provides in polynomial time lower bounds for balancing elephant flows. A methodology based on the rank of the prediction accuracy metrics is applied to compare the hybrid model’s performance with the ARIMA, GARCH, RBF, MLP, and LSTM models. Results show that the FARIMA-RNN model presents lower error rates than the other predictors. Furthermore, we evaluate our proposed heuristic performance on an emulated network with Mininet. The ex periments show that the RDRH solution presents a performance gain compared to the ECMP and Hedera solutions in the round-trip delay and loss metrics in two of the four evaluated scenarios.