Argument Controlled Profiling

Profiling tools relate measurements to code context such as function names in order to guide code optimization. For a more detailed analysis, call path or phase-based profiling enhances the context by call chains or user defined phase names, respectively. In this paper, we propose argument controlled profiling as a new type of context extension using the value of function arguments as part of the context. For a showcase simulation code, we demonstrate that this simplifies and enriches the understanding and analysis of code--in particular recursive functions. Due to the new profiling technique, we found optimizations resulting in more than 16% runtime improvement. Argument controlled profiling is implemented as extension of Callgrind, a simulation-based profiling tool using runtime instrumentation.

• Type: Journal
• Year: 2009
• Where: EUROPAR
• Authors: Tilman Küstner, Josef Weidendorfer, Tobias Weinzierl
• Keywords: Code, Distributed And Parallel Computing, EUROPAR 2009, Particular Recursive Functions, Simulation-based Profiling Tool