Power consumption costs takes upto half of operational expenses of datacenters making power management a critical concern. Advances in processor technology provide fine-grained control over operating frequency and voltage of processors and this control can be used to tradeoff power for performance. Although many power and performance models exist, they have a significant error margin while predicting the performance of memory or file-intensive tasks and HPC applications. Our investigations reveal that the prediction error is due in part to the fact that they do not take frequency AND CPU variations account, rather they just depend on the CPU by itself. In this paper, we empirically derive power and completion time models using linear regression with CPU utilization and operating frequency as parameters. We validate our power model on several Intel and AMD processors by predicting within 2-7% of measured power. We validate our completion time model using five kernels of NASA Parallel Benchmark suite and five CPU, memory and file-intensive benchmarks on four heterogeneous systems and predicting within 1-6% of observed performance. We then show how these models can be employed to realize as much as 15% savings in power while delivering 44% better performance for applications deployed in a virtualized environment.