The following paper got best paper award in NSDI 2023: https://www.usenix.org/conference/nsdi23/presentation/perry.
The authors of this paper "explore a new design point for WAN TE: training a TE decision model on historical data about traffic demands to directly output high-quality TE configurations." The paper presents "the *DOTE (Direct Optimization for Traffic Engineering) *TE framework. DOTE applies stochastic optimization to learn how to map recently observed traffic demands (e.g., empirically-derived traffic demands from the last hour) to the next choice of TE configuration. Using DOTE, providers need only passively monitor traffic to/from datacenters and do not have to onboard applications onto brokers. Directly predicting TE outcomes that optimize TE performance also resolves the objective mismatch between demand prediction and TE performance, yielding TE outcomes that are more robust to traffic unpredictability." Some recent work "speeds up the multicommodity flow computations that underpin TE optimization by effectively breaking the large LPs (*) into smaller problems that can be solved in parallel. However, these approaches still rely on predicted demand matrices. DOTE offers an alternate way to speed up TE: replacing the LP solver with invocations of a fairly small DNN. This has the potential to be innately more efficient." Using DNN DOTE can scale to handle large WAN. Hesham (*) LP is Linear Programming.
_______________________________________________ rtgwg mailing list [email protected] https://www.ietf.org/mailman/listinfo/rtgwg
