With the proliferation of big data initiatives across corporations and academia, the importance of distributed computing frameworks has created many research opportunities related to routing and scheduling the multiple groups of parallel flows generated by MapReduce, Spark, and other similar frameworks. Coflows is the name given to the abstraction of the network flows that distribute the data for parallel computation on distributed nodes and the subsequent redistribution of the results for aggregation or further computation.

The authors propose an algorithm, for both scheduling and routing traffic between nodes in a data center, which aims to minimize the total coflow completion time. Given that each computation stage is dependent on all the data for that stage having been sent and received, minimizing this time is of much greater benefit than the traditional minimization of the average packet transmission time. A single delayed flow will hold up the whole distributed computation.

The authors propose a constraint minimization approach, which being NP hard, they approximate using a binary search over a simplified solution space. Having identified a viable set of paths, the exact route across the network is made with random choices at each hop, such that the constraints are still satisfied. Multiple coflows are handled without rerouting existing traffic, but rather by rescaling the bandwidth utilization along each path so as to accommodate the newly arrived coflow, once again using a binary search. One advantage of the proposed approach is that it does not require upfront knowledge of all future coflows. One criticism is that the provided theoretical performance guarantees are far too wide to be of practical interest.