Our preceding experiments have assumed that resource availability data is collected from every node every 5 minutes, the minimum time granularity of our trace. In a large-scale system, it may be undesirable to collect data about every resource attribute from all nodes that frequently. Thus, we investigate two optimizations that a service placement and migration service might use to reduce measurement overhead. First, the system might simply collect node measurement data less frequently, accepting the tradeoff of reduced accuracy. Second, it might use statistical techniques to predict resource values for one resource based on measurements it has collected of other resource values on the same node, resource values on other nodes, or historical resource values.