A Model-Based Allocator

This section outlines a resource provisioning algorithm that plans least-cost resource slices based on the models from Section 3. The utility OS executive periodically invokes it to adjust the allotments, e.g., based on filtered load and performance observations. The output is an allotment vector for each service, representing a CPU share together with memory and storage allotments $[M, \phi ]$. The provisioning algorithm comprises three primitives designed to act in concert with an assignment planner, which maps the allotted shares onto specific servers and storage units within the utility. The resource provisioning primitives are as follows:

• Candidate plans initial candidate allotment vectors that it predicts will meet SLA response time targets for each service at its load $\lambda$.

• LocalAdjust modifies a candidate vector to adapt to a local resource constraint or surplus exposed during assignment. For example, the assignment planner might place some service on a server with insufficient memory to supply the candidate M; LocalAdjust constructs an alternative vector that meets the targets within the resource constraint, e.g., by increasing $\phi$ to reduce R_S.

• GroupAdjust modifies a set of candidate vectors to adapt to a resource constraint or surplus exposed during assignment. It determines how to reduce or augment the allotments to optimize a global metric, e.g., to minimize predicted average response time. For example, the assignment planner might assign multiple hosted services to share a network storage unit; if the unit is not powerful enough to meet the aggregate resource demand, then GroupAdjust modifies each vector to conform to the constraint.

We have implemented these primitives in a prototype executive for a utility OS. The following subsections discuss each of these primitives in turn.