AppLeS

The focus of the AppLeS (Application-Level Scheduling) project [3] is to design and develop custom application-level scheduling agents for distributed applications. Each application is integrated with its own AppLeS scheduler which develops and implements a custom application schedule, which can adapt to forecasts of deliverable resource performance at execution time. AppLeS agents use performance prediction models, dynamic information, and application-specific information to determine an adaptive custom schedule for their application.

AppLeS schedulers are based on the principle that every potential scheduling decision has a performance impact on the application [2]. An AppLeS agent uses performance models and resource performance forecasts to derive an adaptive custom schedule for its application, and to choose the schedule from among a set of possible candidates that best optimizes the user's performance criteria. To do so, the agent must

• select potentially useful sets of resources,
• derive an application schedule that uses the resources in each set,
• predict the performance an application would attain if it used these resources and schedule, and
• compare different combinations of resources and schedules (based on the performance predicted for each) to choose the one that is likely to yield the best performance.

These functions are implemented by the subsystems depicted in Figure 1.

Figure 1: AppLeS Architecture. Subsystems include the Resource Selector, Schedule Planner and Application Actuator. Information is provided by the user, the system, and by facilities such as the Network Weather Service.