Real-Time Performance Monitoring, Adaptive Control, and Interactive Steering of Computational Grids

The scope of high-performance computing is rapidly expanding from single parallel systems to clusters of heterogeneous sequential and parallel systems. Moreover, as applications become more complex, they grow more irregular, with data-dependent execution behavior, and more dynamic, with time-varying resource demands. Consequently, even small changes in application structure can lead to large changes in observed performance. This performance sensitivity is a direct consequence of the complexity of resource interaction and a clear indication that resource management must evolve with applications. With the emergence of computational grids and their ever-changing resource base, resource management must become more flexible and responsive to changing resource availability and resource demands. To support creation of nimble applications for computational grids, we believe one must not only tightly integrate compilers, languages, libraries, algorithms, problem-solving environments, runtime systems, schedulers, and tools but also eliminate the barrier that separates program creation from execution and post-mortem optimization. This view is based on our experience building performance instrumentation and analysis tools for parallel systems [5, 6, 8], integrating runtime measurement and deep compile-time analysis [1], and from an ongoing series of discussions application, compiler, library, and runtime system developers. In this integrated model, high-level problem solving environments (PSEs) allow users to compose programs from configurable modules, each with a performance specification interface that describes the expected module behavior as a function of available resources. Using PSE specifications, whole program compilers generate configurable object programs that include embedded performance instrumentation and support for runtime validation of performance specifications. During execution, these configurable programs dynamically adapt to changing resource availability, invoking compilers to generate code variants optimized for current conditions and exploiting adaptive runtime systems to negotiate program behavior and resource management. To support dynamic performance adaptation and distributed optimization in the grid environment, we are building a suite of performance instrumentation, analysis and presentation tools that includes distributed performance sensors and resource policy actuators, fuzzy logic rule bases for adaptive control, and immersive visualization systems for real-time visualization and direct manipulation of software behavior. With this context, the remainder of this paper is organized as follows. In §2, we