Appendix

File format of statistics files

Statistic files (for an example see ) are simply text files which contain the necessary data. The first line is always ignored but should look similar to that in the example as it simplifies the understanding for the human reader. All values in a statistic file are simply separated by an arbitrary number of spaces.

PatternName  MetricID Count  Mean   Median  Minimum  Maximum   Sum Variance Quartil25 Quartil75
LateBroadcast       6     4 0.010 0.000031 0.000004 0.042856 0.042 0.000459
- cnode: 5 enter: 0.245877 exit: 0.256608 duration: 0.042856

WaitAtBarrier      18    20 0.018 0.006477 0.000002 0.065293 0.369 0.000698  0.000040  0.047409
- cnode: 14 enter: 0.192332 exit: 0.192378 duration: 0.000100
- cnode: 12 enter: 0.326120 exit: 0.335651 duration: 0.065293

BarrierCompletion  17    20 0.000 0.000005 0.000002 0.000018 0.000 0.000000  0.000003  0.000009
- cnode: 14 enter: 0.192332 exit: 0.192378 duration: 0.000009
- cnode: 12 enter: 0.159321 exit: 0.165005 duration: 0.000018

WaitAtIBarrier     27   144 0.001 0.000027 0.000001 0.028451 0.212 0.000028  0.000002  0.000437
- cnode: 11 enter: 0.297292 exit: 0.297316 duration: 0.000057
- cnode: 10 enter: 0.322577 exit: 0.332093 duration: 0.028451

For each pattern there is a line which contains at least the pattern name (as plain text without spaces), its corresponding metric id in the CUBE file (integer as text) and the count – i.e., how many instances of the pattern exist (also as integer). If more values are provided, there have to be the mean value, median, minimum and maximum as well as the sum (all as floating point numbers in arbitrary format). If one of these values is provided, all have to. The next optional value is the variance (also as a floating point number). The last two optional values of which both or none have to be provided are the 25% and the 75% quantile, also as floating point numbers.

If any of these values is omitted, all following values have to be omitted, too. If for example the variance is not provided, the lower and the upper quartile must not be provided either.

In the subsequent lines (there can be an arbitrary number), the information of the most severe instances is provided. Each of these lines has to begin with a minus sign (-). Then the text cnode:, followed by the cnode id of this instance in the CUBE file (integer as text) is provided. The same holds for enter, exit and duration (floats as text).

The begin of the next pattern is indicated by a blank line.

Message Passing Interface Forum: MPI: A Message Passing Interface Standard, June,1995, http://www.mpi-forum.org

OpenMP Architecture Review Board: OpenMP Fortran Application Program Interface — Version 2.5, May,2000 http://www.openmp.org

K. L. Karavanic and B.Miller, A Framework for Multi-Execution Performance Tuning, Parallel and Distributed Computing Practices, 4(3), 2001, September

F.Song and F.Wolf and N.Bhatia and J.Dongarra and S.Moore, An Algebra for Cross-Experiment Performance Analysis, Proc. of ICPP 2004, 63-72, 2004, Aughust, Montreal, Canada

F.Wolf and B.Mohr and J.Dongarra and S.Moore, Efficient Pattern Search in Large Traces through Successive Refinement, Proc. of the European Conference on Parallel Computing (Euro-Par), August - September, 2004 Lecture Notes in Computer Science, Springer,Pisa, Italy,

J.Labarta and S.Girona and V.Pillet and T.Cortes and L.Gregoris, DiP: A Parallel Program Development Environment, Proc. of the 2nd International Euro-Par Conference, Springer, 665-674 Lyon, France, August, 1996

Barcelona Supercomputing Center, Paraver: Obtain Detailed Information from Raw Performance Traces,Oct,2008, http://www.bsc.es/plantillaA.php?cat_id=485

H.Brunst and W.E.Nagel, Scalable Performance Analysis of Parallel Systems: Concepts and Experiences Proc. of the Parallel Computing Conference (ParCo), 2003, Dresden, Germany

Technical University Dresden, Vampir - Performance Optimization, Oct, 2008 http://vampir.eu/

World Wide Web Consortium, Extensible Markup Language (XML) 1.0 (Second Edition), October, 2000 http://www.w3.org/TR/REC-xml

Sameer S. Shende and Allen D. Malony, The TAU Parallel Performance System, International Journal of High Performance Computing Applications,20(2), 287–331 SAGE Publications, Summer, 2006

The Scalasca Development Team scala.nosp@m.sca@.nosp@m.fz-ju.nosp@m.elic.nosp@m.h.de, Cube 4.2.0- Cube Derived Metrics, Usage and syntax documentation


Scalasca