Scalable Performance Analysis of Large-Scale Parallel Applications
Brian Wylie
John von Neumann Institute for Computing (NIC) Forschungszentrum Jülich GmbH, D-52425 Jülich, Germany
October 12, 2006
10:00 AM
ORNL 5100-Auditorium
Host: Phil Roth (rothpc@ornl.gov)
ABSTRACT:
The SCALASCA project is developing a new generation of tools for
scalable performance analysis of large-scale parallel applications,
building on extensive experience gained with the KOJAK toolset developed
by Forschungszentrum Juelich and the University of Tennessee. �First
steps have extended KOJAK to significantly improve the scalability of
trace capture and analysis of thousands of MPI processes, by avoiding
trace re-writing and exploiting a replay-based parallel analysis. �As
demonstrated with Top-10 IBM Blue Gene/L and Cray XT3 systems,
previously impractical performance analysis has now been enabled.
# # #