The VampirTrace plugin counter interface: Introduction and examples

Abstract

The growing complexity of microprocessors is not only driven by the current trend towards multi-core architectures, but also by new features like instruction set extensions, additional function units or specialized processing cores. The demand for more performance and scalability also results in an increasing complexity of the software stack: operating systems, libraries, and applications all need to exploit more parallelism and new functionalities in order to meet this demand. Both aspects - hardware and software - put pressure on performance monitoring infrastructures that face two conflictive requirements. On the one hand, performance tools need to be somewhat stable without entailing significant software changes with every additional functionality. On the other hand they need to be able to monitor the influence of new hardware and software features. We therefore present a plugin interface for our performance monitoring software VampirTrace that allows users to write libraries that feed VampirTrace with data from new (platform dependent) performance counters as well as hardware features that may not be accessed by Open Source software. This paper describes the interface in detail, analyzes its strength and weaknesses, depicts examples, and provides a comparison to other plugin-like performance analysis tools. © 2011 Springer-Verlag Berlin Heidelberg.