Traffic Engineering (TE) typically shifts flows to different paths to optimize resource utilization in the network. These shifts can cause out-of-order packets or unnecessary retransmissions, resulting in a performance deterioration for flows. Minimizing this by keeping paths persistent for a majority of flows may reduce the overall impact of flow shifts. We investigate different shifting strategies that are subject to a fundamental tradeoff between reducing the number of shifted flows and fast convergence times, i.e., the time to reach the updated TE configuration. We developed the flow-level simulation tool-chain TEFlowSim to evaluate the shifting strategies based on real-world traffic. We find that keeping most flows on the same path while shifting only high-bandwidth, long-running elephant flows reduces the number of shifted flows to a minimum while fulfilling the targeted shift of bandwidth. Our experiments show that compared to state-of-the-art hash-based strategies, only shifting elephant flows reduces the number of shifted flows by up to 45x while achieving fast convergence.
Schichtholz et al. (Thu,) studied this question.