An Extensible Timing Infrastructure for Adaptive Large-scale Applications

TL;DR

This paper presents an extensible timing infrastructure within the Cactus Framework that provides accurate, reliable timing data for large-scale scientific applications, enabling profiling and dynamic adaptation during runtime.

Contribution

It introduces a flexible, well-designed timing API for Cactus that integrates multiple timing sources and demonstrates its use in adaptive scientific applications.

Findings

Provides access to multiple timing sources including hardware clocks and high-level interfaces

Enables applications to profile and adapt dynamically based on timing data

Demonstrates practical use in a scientific application for runtime adaptation

Abstract

Real-time access to accurate and reliable timing information is necessary to profile scientific applications, and crucial as simulations become increasingly complex, adaptive, and large-scale. The Cactus Framework provides flexible and extensible capabilities for timing information through a well designed infrastructure and timing API. Applications built with Cactus automatically gain access to built-in timers, such as gettimeofday and getrusage, system-specific hardware clocks, and high-level interfaces such as PAPI. We describe the Cactus timer interface, its motivation, and its implementation. We then demonstrate how this timing information can be used by an example scientific application to profile itself, and to dynamically adapt itself to a changing environment at run time.