Predicting SPEC benchmarks values for untested systems

SPEC benchmarks are often used to compare the relative performance of servers. Typical areas of applications are datacenter consolidation and what-if analysis. Unfortunately, each server comes with different configurations (e.g., number of processors, memory size) while the published SPEC results are available only for a small subset of the configurations, typically the most powerful. The problem we are trying to solve in this work is how to scale down, or up, a SPEC result in order to account for different number of processors with respect to the benchmarked configuration.