Resource Management in Multi-* Clusters: Cloud Provisioning

Cloud computing – running large-scale computationand data-intensive services on inexpensive onthe-fly clusters – has become increasingly popular in recent years due to the advent of MapReduce and similar large-scale data parallel systems. Cloud provisioning, i.e., allocating resources for cluster requests, is the first step toward instantiating such clusters. In this report, we formally define the cloud provisioning problem (CPP) and show it to beNP-hard. We present a flow network-based model of CPP and describe a generalized framework for optimally solving variants of CPP (e.g., CPP with load balancing (LBCPP), fault tolerance (K-FTCPP) etc.) using linear and non-linear mixed integer programs. Since mixed integer programs are generallyNP-hard, we also sketch a linear programming relaxation and randomized rounding-based approximation algorithm for 1-FTCPP (a specialized/simpler variant of K-FTCPP) and propose a rounding-based heuristic for the problem.