Joseph S. Barrera

Modular real-time resource management in the Rialto operating system

This paper describes ongoing investigations into algorithms for modular distributed real-time resource management. These investigations are being conducted in the context of the Rialto operating system-an object-based real-time kernel and programming environment currently being developed within Microsoft Research. Some of the goals of this research include developing appropriate real-time programming abstractions to allow multiple independent real-time programs to dynamically coexist and share resources on the same hardware platforms. Use of these abstractions is intended both to allow individual applications to reason about their own resource requirements and for per-machine system resource planner applications to reason about and control resource allocations between potentially competing applications. The set of resources being managed is dynamically extensible, and may include remote resources in distributed environments. The local planner conducts resource negotiations with individual applications on behalf of the user, with the goal of maximizing the users perceived utility of the set of running applications with respect to resource allocations for those applications.

Self-tuning systems software

Systems software that tunes and reconfigures itself is both feasible and increasingly necessary. The necessity of self-tuning systems arises from the increased complexity of systems software combined with the broader audience for such systems. We demonstrate the feasibility of self-tuning systems by constructing one based on an architecture which separates the tasks of defining expectations, measuring actual performance, analyzing measurements in comparison with expectations, and performing actions in response to analysis which can range from gathering more data to reconfiguring major components of the system. This separation of responsibilities allows systems components to concentrate on performing well within narrower bands of operation, leaving the analysis agent to make more global and longer term decisions regarding the best operating parameters and component implementations to use.<<ETX>>

Invocation chaining: manipulating lightweight objects across heavyweight boundaries

Invocation batching combines multiple object invocations into a single message; result chaining makes results from one batched invocation available to the other invocations batched with it. Invocation chaining, or the combination of invocation batching with result chaining, is the key to allowing lightweight objects to be manipulated efficiently across heavyweight boundaries, whether between machines, between address spaces, or between user and kernel. By reducing the number of boundary crossings, invocation chaining reduces the total cost of invocation, making it more effective than previous solutions such as asynchronous messaging. This paper describes an initial implementation of invocation chaining.<<ETX>>