Comparison of CP and IP Techniques for Data Transfer Planning

Planning of data transfers and allocation of network resources becomes an interesting problem when bandwidth of a requested traffic is comparable to a capacity of available networks. An interactive high-quality media distribution for collaborative environments is one of the important applications of the problem. The CoUniverse framework (Liška and Holub 2008) was proposed to handle these applications automatically for end-users. A planning module is one of the crucial components of the framework. Each demand is represented by a transfer of media stream from a producer to a set of consumers. To plan each demand, a proper distribution of the stream through a set of distributors is computed. There is no temporal reasoning since all allocated resources must be available all the time during the interaction. To allow for real-time experience with the CoUniverse, it is necessary to compute a plan within a few seconds. Fast planner responses also allow a graceful reaction to resource failures or changes such as link outage. More precisely, the CoUniverse monitors the network and can process replanning from scratch if needed. Variety of network optimization problems were described in the survey published by Simonis (2006). The problem we are interested in is related to a path placement problem where the decision about allocating path for a demand should be done. Since each demand corresponds to a data transfer from a producer to a set of consumers, a communication graph is a tree with the producer in its root, consumers at leafs and media distributors at internal nodes. Therefore, we call the problem tree placement. The link-based model for this problem was described in (Holub, Rudová, and Liška 2010) using techniques of constraint programming (CP). Java constraint solver Choco 2.1.11 was applied to implement the model. The constraint programming solver was improved over the years since its first implementation (Liška and Holub 2008) and allows solving of medium-sized data instances currently. A new integer programming (IP) solver was proposed in (Troubil and Rudová 2010) and utilizes the Gurobi Optimizer 3.0.22. This solver was implemented in