Automatic Performance Analysis of Parallel Applications with KOJAK

Colloq: Speaker: 
Felix Wolf
Colloq: Speaker Institution: 
University of Tennessee, Knoxville
Colloq: Date and Time: 
Wed, 2004-10-06 14:00
Colloq: Location: 
ORNL 5700-L204
Colloq: Host: 
Jeffrey S. Vetter
Colloq: Host Email:
Colloq: Abstract: 
KOJAK is an integrated performance evaluation environment for MPI and/or OpenMP applications. KOJAK generates event traces from running applications and automatically searches them offline for execution patterns indicating inefficient performance behavior. Its main feature is the ability to automatically diagnose the reasons for inefficient runtime behavior on a high level of abstraction. The KOJAK project was started in 1998 at Forschungszentrum Juelich. In 2003, it was turned into a transatlantic collaboration between Forschungszentrum Juelich and the University of Tennessee. This talk introduces KOJAK, describes two recent extensions done in Tennessee, and gives an outlook on future work. The first extension is an optimization of the search algorithm taking advantage of specialization relationships between different patterns and leading to a significant runtime improvement as well as to more compact pattern specifications. The second extension is a performance algebra that supports multi-experiment analysis by providing operators to combine the data from different experiments and view the results as a single \derived\experiment. Using the same representation for derived experiments and original experiments provides access to the derived behavior using familiar metaphors and tools in addition to an arbitrary and easy composition of operations.
Colloq: Speaker Bio: