Performance-Monitoring Counters Library, for Intel/AMD Processors and Linux
This example introduces
   rabbit command-line options -output output_directory -Raw n

Return to Main Menu


Command-Line Options -output output_directory rabbit usually sends its performance summary to stderr. This option will create an output directory (if it does not already exist) to hold the results of an experiment. The directories ., .., `pwd` are not accepted [too messy]. rabbit will overwrite the contents of the directory if file names agree; this means that some old files could remain. Five classes of files are written: an input record (filename, input), as with the -Input option. an output summary (filename, summary), more than would normally be produced to stderr if the -output option had not been used. data files (filenames, data.0, data.1, ...), one event pair per file, one line per sample. The -flush option controls how often the internal buffers are written to the data files. The -sample option controls the sampling rate. The -rotate option controls the rate of change of events. The -trim option can omit the first few samples. raw data files (filenames rawdata.0, rawdata.1, ...), one event pair per file, one or more lines per sample. a gnuplot script (filename, plot) to display the data files. -Raw [0|1|2] This option affects the format of the raw data files. It is similar to the -Stats option.
Examples rabbit foo # rabbit's summary follows foo's stdout rabbit foo 2> foo.perf # rabbit's summary is redirected [bash] rabbit foo > foo.out # foo's stdout is redirected rabbit -o z foo # foo's stdout is not affected gnuplot z/plot # view the performance data
Notes If you have a file system that is cross-mounted from several machines, you can run 'rabbit -o z foo' on one system, and 'gnuplot z/plot' on another, for a more-or-less real-time monitor. The gnuplot interactive command 'load "z/plot"' also works, and can be used repeatedly to update the graphs without changes to the windows. If a data file is produced with 0 events/sec for all samples of both counters, then gnuplot will complain, as a result of using 'set logscale y' in the plot script. The raw data files contain the elapsed time of a measurement since the start of the experiment, and the number of cycles and events counted in the time interval ending at this time. The data files have the same information, and averages, as events per second. The format of the raw data files is affected by the command-line option -Raw [0|1|2]; more detailed information can be provided.

Performance-Monitoring Counters Library, for Intel/AMD Processors and Linux
Author: Don Heller, dheller@scl.ameslab.gov
Last revised: 22 November 2000