What is the best strategy to optimize costs while maintaining the performance of a business-critical workload running on a fixed set of Compute Engine instances for several months?