Designing Controllable Computer Systems

Adaptive control theory is emerging as a viable approach for the design of self-managed computer systems. This paper argues that the systems community should not be concerned with the design of adaptive controllers—there are off-the-shelf controllers that can be used to tune any system that abides by certain properties. Systems research should instead be focusing on the open problem of building and configuring systems that are amenable to dynamic, feedback-based control. Currently, there is no systematic approach for doing this. To that aim, this paper introduces a set of properties derived from control theory that controllable computer systems should satisfy. We discuss the intuition behind these properties and the challenges to be addressed by system designers trying to enforce them. For the discussion, we use two examples of management problems: 1) a dynamically controlled scheduler that enforces performance goals in a 3-tier system; 2) controlling the number of blades assigned to a workload to meet performance goals within power budgets.