HomeCollecting Trace Data for HPC Applications
Collecting Trace Data for HPC Applications
Date:Wednesday, May 27, 2009
Time:1:00pm - 2:00pm
Location/:LBNL Bldg. 50F, Room 1647
Speaker:
Noel Keen
Computational Research Division
Lawrence Berkeley National Laboratory
Abstract:
By intercepting calls to the system LIBC library, trace data was
collected, specifically I/O and memory allocation operations.
Functionality was extended to allow IPM (light-weight MPI profiling
tool) to
treat these intercepted LIBC calls in much the same way that IPM
handles
MPI calls. Intercepting calls to LIBC calls only requires that the
application
code be re-linked to a static library. A tool was written to
post-process
and visualize the trace data that can be used as an aid to
performance profiling.
Data was collected on Franklin (Cray XT at NERSC) using several
codes, including
I/O intensive applications (MADbench and Chombo-IO), and an
astrophysics
Particle-in-Cell (PIC) code named Charm that uses Chombo.
Host of Seminar:
Tony Drummond