Integrating Job Management and Machine Management in Google Data Centers

In cloud based services, many jobs run for months. This is a time frame that is on the order of machine lifetimes. As a result, scheduling the administration of machines (e.g., repairs, adding new machines, draining machines) must be integrated with scheduling jobs. At Google, we approach this problem by using an abstraction called shapes to describe both machines and jobs. I will discuss the shapes abstraction, and I will also describe optimization problems that we solve using shapes.