The Efficient Characterization of Adaptive Applications through Fine Profiling based on Data Locality

Q. Liu and K.A. Tomko (USA)


Parallel computing, mesh-based, adaptive application, profiling


The efficient characterization of adaptive parallel applications is usually challenging due to their complexity and large problem size. Unlike traditional profiling approaches which target the tracing of events or determining performance parameters for subroutines, the approach described in this paper attempts to discover the inherent adaptivity of parallel applications mapped to the computation domain/mesh, which are independent of runtime environment, so as to aid in the performance tuning of parallel applications, especially dynamic load balancing and repartitioning. Our profiling scheme only requires one-time execution of the target program on any platform to generate a sequence of traces with timestamps. The traces can then be fed to simulations under various system configurations while independent of the real application. Preliminary experiments have been performed to evaluate the proposed profiling techniques.

Important Links:

Go Back