In grid or cloud computing, the optimal location and capacity of data centers (hosting the servers executing the remote users' tasks) depends on the available network resources, esp. link bandwidths providing connectivity to the candidate data center locations. In this study, we propose to jointly optimize the link dimensioning and the location of the servers in an optical grid or cloud, where we exploit the anycast principle for resiliency against either link or node failures. (The latter implies that we will allow to use a different server location under failure conditions.) While the problem has some resemblance with either the classical p-center or k-means location problems, the resilience requirement exploiting anycast significantly increases complexity, since we need link/node disjoint paths to possibly two different locations for each request.
We propose three different decomposition schemes, each resulting in a different model, to address scalability issues encountered in traditional optimization models. We extensively compare the models with computational experiments. The first model is quite efficient and scalable, under the assumption of several backup paths for a given source node, i.e., more complexity at the control network level. If we only allow a single backup path to be followed for all requests from a given source node, then the second model is the most efficient. Our third model is a direct extension of the model proposed for the p-median problem, i.e., the closest classical problem of operation research, but it is the least efficient and scalable one.
Published May 2014 , 34 pages