Minimum congestion mapping in a cloud

We study a basic resource allocation problem that arises in cloud computing environments. The physical network of the cloud is represented as a graph with vertices denoting servers and edges corresponding to communication links. A workload is a set of processes with processing requirements and mutual communication requirements. The workloads arrive and depart over time, and the resource allocator must map each workload upon arrival to the physical network. We consider the objective of minimizing the congestion. We show that solving a subproblem about mapping a single workload to the physical graph essentially suffices to solve the general problem. In particular, an a-approximation for this single mapping problem gives an O(a log nD)-competitive algorithm for the general problem, where n is the number of nodes in the physical network and D is the maximum to minimum workload duration ratio. We also show how to solve the single mapping problem for two natural class of workloads, namely depth-d-trees and complete-graph workloads. For depth-d tree, we give an nO(d) time O(d2 log (nd))-approximation based on a strong LP relaxation inspired by the Sherali-Adams hierarchy.