#!/usr/bin/env python # # pprof.py, v0.5: Create performance profiles. # Written by Michael P. Friedlander # # $Revision: 1.11 $ $Date: 2003-04-24 12:11:21-05 $ # # I'm still learning Python, so bear with me! # import getopt, sys, re import Gnuplot, Gnuplot.funcutils from MA import * from string import atoi, atof PROGNAME = "pprof.py" def usage(): instructions = """ Usage: %s [OPTION]... [FILE 1] [FILE 2]...[FILE N] Create a performance profile chart from metrics in FILEs. Output is an eps file sent to stdout. Example 1: Profile the metrics in column 3 of the files solver1, solver2, and solver3. Use a log2 scale for the x-axis. Redirect the stdout to profile.eps. %s -l 2 -c 3 solver1 solver2 solver3 > profile.eps Example 2: Specify a title, linestyle and failure threshold. Pop up an X window. %s -c 3 -t "Plot title" --linestyle "linespoints" \\ --term "x11" solver1 solver2 solver3 See Dolan and More', "Benchmarking optimization software with performance profiles", available at http://www-unix.mcs.anl.gov/~more/cops/ Options -c, --column=COLUMN get metrics from column COLUMN (default 1) -h, --help get some help -l, --log=BASE logBASE scale for x-axis (default linear) --legend insert a legend --linestyle=STYLE use STYLE as Gnuplot line style (default steps) --sep RE use regexp RE to indicate new column (default space) --term=TERM use TERM Gnuplot terminal (default postscript) -t, --title=LABEL use LABEL as title (default none) -x, --xlabel=LABEL use LABEL for x-axis (default none) -y, --ylabel=LABEL use LABEL for y-axis (default none) - Use non-positive values to indicate that the algorithm failed. - Use --sep 'S' to indicate columns are separated by character S. - Use --sep "r'RE'" to separate instead by a regular expression. - Any line starting with a %% or a # is ignored. """ % (PROGNAME,PROGNAME,PROGNAME) print instructions sys.exit(0) def commandline_err(msg): sys.stderr.write("%s: %s\n" % (PROGNAME, msg)) sys.stderr.write("Try '%s --help' for more information.\n" % PROGNAME) sys.exit(1) class OptionClass: def __init__(self): self.datacol = 1 self.legend = None self.linestyl = 'steps lw 4' self.logscale = None self.sep = '\s+' # The GnuPlot postscript driver seems most sophisticated. # Also, probably the easiest format for LaTeX to deal with. self.term = 'postscript eps color solid' self.title = None self.xlabel = None self.ylabel = None def parse_cmdline(arglist): """Parse argument list""" if len(arglist) == 0: usage() options = OptionClass() try: optlist, files = getopt.getopt(arglist, 'hl:c:x:y:t:', ["column=", "help", "legend", "linestyle=", "sep=", "log=", "term=", "title=", "xlabel=", "ylabel=", ]) except getopt.error, e: commandline_err("%s" % str(e)) if len(files) < 2: usage() for opt, arg in optlist: if opt in ("-c", "--column"): options.datacol = atoi(arg) elif opt in ("-h", "--help"): usage() elif opt == "--linestyle" : options.linestl = arg elif opt == "--legend" : options.legend = 1 elif opt in ("-l", "--log"): options.logscale = atof(arg) elif opt == "--sep" : options.sep = arg elif opt == "--term" : options.term = arg elif opt in ("-t", "--title"): options.title = arg elif opt in ("-x", "--xlabel"): options.xlabel = arg elif opt in ("-y", "--ylabel"): options.ylabel = arg return (options, files) class MetricsClass: def __init__(self, solvers, opts): self.metric = None self.nprobs = [] self.perf = [] self.solvers = solvers self.nsolvs = len(solvers) self.sep = opts.sep map(self.add_solver, solvers) def add_solver(self, fname): # Reg exp: Any line starting (ignoring white-space) # starting with a comment character. Also col sep. comment = re.compile(r'^[\s]*[%#]') column = re.compile(self.sep) # Grab the column from the file metrics = [] file = open(fname, 'r') for line in file.readlines(): if not comment.match(line): line = line.strip() cols = column.split( line ) data = atof(cols[opts.datacol - 1]) metrics.append(data) file.close() if self.metric is not None: self.metric = concatenate((self.metric, [metrics])) else: self.metric = array([metrics]) # Current num of probs grabbed nprobs = len(metrics) if not self.nprobs: self.nprobs = nprobs elif self.nprobs <> nprobs: commandline_error("All files must have same num of problems.") def prob_mets(self, prob): return masked_less_equal(self.metric[:,prob], 0) class RatioClass: def __init__(self, MetricTable): # Create empty ratio table nprobs = MetricTable.nprobs nsolvs = MetricTable.nsolvs self.ratios = 1.0 * zeros((nprobs+1, nsolvs)) # Compute best relative performance ratios across # solvers for each problem for prob in range(nprobs): metrics = MetricTable.prob_mets(prob) best_met = minimum(metrics) self.ratios[prob+1,:] = metrics * (1.0 / best_met) # Sort each solvers performance ratios for solv in range(nsolvs): self.ratios[:,solv] = sort(self.ratios[:,solv]) # Compute largest ratio and use to replace failures entries self.maxrat = maximum(self.ratios) self.ratios = filled(self.ratios, 1.01 * self.maxrat) def solv_ratios(self, solver): return self.ratios[:,s] ################ # Main program # ################ opts, solvers = parse_cmdline(sys.argv[1:]) metrics = MetricsClass(solvers, opts) pprofs = RatioClass(metrics) nprobs = metrics.nprobs nsolvs = metrics.nsolvs # Create a performance profile graph persist_val = ( opts.term in ('X11', 'x11') ) g = Gnuplot.Gnuplot(persist=persist_val, debug=0) # Generate the x-axis data ydata = arange(nprobs+1) * (1.0 / nprobs) # Set the x-axis ranges maxrat = pprofs.maxrat + 1.0e-3 if opts.logscale: g('set logscale x %f' % opts.logscale) minrat = 1 else: minrat = 0 g('set xrange [%f:%f]' % (minrat, maxrat / 1.00)) # The y-axis range is fixed g('set yrange [0:1]') # Set graph properties if opts.legend: g('set key bottom right') else: g('set nokey') if opts.term: g('set term ' + opts.term) if opts.title: g.title(opts.title) if opts.xlabel: g.xlabel(opts.xlabel) if opts.ylabel: g.ylabel(opts.ylabel) # Generate arguments for the gplot command plotargs = [] for s in range(nsolvs): sname = solvers[s] srats = pprofs.solv_ratios(s) plotargs.append(Gnuplot.Data(srats, ydata, title=sname, inline=0, with=opts.linestyl)) # Create the plot apply(g.plot, plotargs) # This fixes a Gnuplot bug. (Thanks to Matt Knepley.) g.gnuplot.gnuplot.close()