Scaling the Web Mapping Service-Level Agreements in Distributed Applications

In a previous column, I discussed how to builddistributed systems from quality-of-service(QoS)-aware software components.1 I also described a design by which individual compo-nents can engage in QoS negotiation and per-form admission control so that new incoming requests don’t violate committed QoS require-ments.2 But what happens in a distributed application made up of several QoS-aware com-ponents? Moreover, what if the application has a global service-level agreement (SLA) for max-imum end-to-end execution time? The problem I discuss here is how to determine which SLAs should be negotiated at the component level so that the global SLA is satisfied at the minimum possible cost. A Model for the Application Let’s say that a distributed application is made up of M software components. Each component m (m = 1, …, M) is visited, on average, Vm times during the application’s execution. When there is no congestion at component m — that is, when only one request is executing on it — the execution time of the application per visit to the component has the minimum possible value. Let Emmin be this value. Let E be the appli-cation’s execution-time SLA, and let Emmax be the execution-time SLA per visit to component m: clearly, Emmax ≥ Emmin. The relationship between the application SLA and th