A Plugin Architecture for the TAU Performance System

摘要: Several robust performance systems have been created for parallel machines with the ability to observe diverse aspects of application execution on different hardware platforms. All these are designed objective support measurement methods that efficient, portable, and scalable. For reasons, infrastructure is tightly embedded code runtime environment. As software evolve, especially towards more heterogeneous, asynchronous, dynamic operation, it expected requirements observation awareness will change. instance, heterogeneous introduce new types data capture behaviors characterize. Furthermore, there a growing interest in interacting situ analytics policy-based control. The problem an existing system architecture could be constrained its evolve meet requirements. paper reports our research efforts address this concern context TAU Performance System. In particular, we consider use powerful plugin model both capabilities extend functionality ways was not necessarily conceived originally. supports three paradigms: EVENT, TRIGGER, AGENT. We demonstrate how each operates under several scenarios. Results from larger-scale experiments shown highlight fact efficiency robustness can maintained, while flexibility programmability offered leverages power core allowing significant compelling extensions realized.