ft-compute-stats

#!/usr/bin/env python

from __future__ import division

import numpy

import optparse
import sys
import os

from math import ceil

from os.path import splitext

import itertools as it

def decode_key_value_filename(name):
    "Map key=value_otherkey=other-value names to proper dictionary."
    params = {}
    parts = name.split('_')
    for p in parts:
        kv = p.split('=')
        k = kv[0]
        v = kv[1] if len(kv) > 1 else None
        params[k] = v
    return params

def print_cols(cols, first_row_prefix='# ', other_rows_prefix='  '):
    col_widths = [max((len(str(x)) for x in col)) for col in cols]

    prefs = it.chain([first_row_prefix], it.repeat(other_rows_prefix))

    for prefix, row in it.izip(prefs, range(0, max((len(c) for c in cols)))):
        reached_end = True
        data = [col[row] if row < len(col) else '' for col in cols]
        print '%s%s' % (prefix,
                        ", ".join([' ' * (c - len(str(f))) + str(f)
                                   for (c, f) in it.izip(col_widths, data)]))


def print_rows(rows, first_row_prefix='# ', other_rows_prefix='  '):
    col = 0
    col_widths = []
    field_widths = True
    while field_widths:
        field_widths = [len(str(row[col])) for row in rows if len(row) > col]
        if field_widths:
            col_widths.append(max(field_widths))
        col += 1

    prefs = it.chain([first_row_prefix], it.repeat(other_rows_prefix))

    for prefix, row in it.izip(prefs, rows):
        print '%s%s' % (prefix,
                        ", ".join([' ' * (c - len(str(f))) + str(f)
                                   for (c, f) in it.izip(col_widths, row)]))



def stats_for_file(fname, scale):
    n    = 0
    max  = 0
    p95  = 0
    p99  = 0
    p999 = 0
    min  = 0
    med  = 0
    avg  = 0
    std  = 0
    var  = 0

    size = os.stat(fname).st_size
    if size:
        samples = numpy.memmap(fname, dtype='float32', mode='c')

        n = len(samples)
        if n > 0:
            samples *= scale
            max  = numpy.amax(samples)
            p95  = numpy.percentile(samples, 95.0)
            p99  = numpy.percentile(samples, 99.0)
            p999 = numpy.percentile(samples, 99.9)
            med  = numpy.median(samples)
            avg  = numpy.mean(samples)
            min  = numpy.amin(samples)

            std = numpy.std(samples, ddof=1)
            var = numpy.var(samples)

    return [n, max, p999, p99, p95, avg, med, min, std, var]

o = optparse.make_option

opts = [
    o('-p', '--cycles-per-usec', action='store', dest='cycles', type='float',
      help='how many cycles per usec'),
    o(None, '--hist', action='store_true', dest='want_hist',
      help='generate a histogram'),
    o('-b', '--bin-size', action='store', dest='bin_size', type='float',
      help='size of each bin in histogram'),
    o('-n', '--normalize-counts', action='store_true', dest='normalize',
      help='give relative frequency, not absolute sample counts'),
    o('-c', '--cumulative', action='store_true', dest='cumulative',
      help='report cumulative counts (i.e., CDF)'),
    o(None, '--percentiles', action='store_true', dest='want_percentiles',
      help='produce table of percentiles'),
    o('-r', '--resolution', action='store', dest='resolution', type='float',
      help='set resolution of percentiles table'),
    o(None, '--percent', action='store_true', dest='want_percent',
      help='give relative frequency as a percentage'),
]

defaults = {
    'cycles'    : None,
    'want_hist' : False,
    'bin_size'  : 1000,
    'normalize' : False,
    'want_percent' : False,
    'cumulative' : False,
    'want_percentiles' : False,
    'resolution' : 0.1,
}

options = None

def to_str(x):
    if type(x) == str:
        return x
    if type(x) == int:
        return "%d" % x
    else:
        return "%.5f" % x

STATS_HEADERS = [
    "Plugin", "#cores", "Overhead", 'Unit', 'Scale', "#tasks",
    "#samples",
    "max", "99.9th perc.", "99th perc.", "95th perc.",
    "avg", "med", "min", "std", "var",
    "file"
]

def get_stats(fname):
    name, ext = splitext(fname)
    conf = decode_key_value_filename(name)

    if 'overhead' in conf and conf['overhead'].rfind('-LATENCY') != -1:
        # latency is stored in nanoseconds, not cycles
        scale = 1 / 1000 # convert from nanoseconds
        unit = 'microseconds'
        scale_desc = '1/1000.00'
    elif options.cycles is None:
        scale = 1
        unit = 'cycles'
        scale_desc = '1'
    else:
        # convert from cycles to usec
        scale = 1 / options.cycles
        unit = 'microseconds'
        scale_desc = '1/%.2f' % options.cycles

    stats = stats_for_file(fname, scale)
    if 'locks' in conf:
        sched = '%s_locks=%s' % (conf['scheduler'], conf['locks'])
    elif 'scheduler' in conf:
        sched = conf['scheduler']
    else:
        sched = 'UNKNOWN'

    ohead = conf['overhead'] if 'overhead' in conf else 'UNKNOWN'
    n = conf['n'] if 'n' in conf else '*'
    m = conf['m'] if 'm' in conf else '*'

    info = [sched, m, ohead, unit, scale_desc, n]
    finfo = [fname]
    return [to_str(x) for x in  info + stats + finfo]

def make_bins(max_val):
    num_bins = int(ceil(max_val / options.bin_size)) + 1
    return [x * options.bin_size for x in range(0, num_bins)]

def iter_percentiles(resolution):
    percentile = 0.0
    while percentile < 100:
        yield percentile
        percentile += resolution
    yield 100.0

def make_percentiles(resolution):
    return list(iter_percentiles(resolution))

def hist_file(fname, scale):
    max_val = 0
    hist = []

    size = os.stat(fname).st_size
    if size:
        samples = numpy.memmap(fname, dtype='float32', mode='c')

        n = len(samples)
        if n > 0:
            samples *= scale
            max_val = numpy.amax(samples)
            bins = make_bins(max_val)
            hist, _ = numpy.histogram(samples, bins)
            if options.cumulative:
                hist = numpy.cumsum(hist)
            if options.normalize:
                hist = [h/n * (100 if options.want_percent else 1) for h in hist]

    return (max_val, hist)


def percentiles_file(fname, percentiles):
    percs = []

    size = os.stat(fname).st_size
    if size:
        samples = numpy.memmap(fname, dtype='float32', mode='c')

        n = len(samples)
        if n > 0:
            percs = numpy.percentile(samples, percentiles)

    return percs


if __name__ == '__main__':
    # FIXME: would be nicer with argparse
    parser = optparse.OptionParser(option_list=opts)
    parser.set_defaults(**defaults)
    (options, files) = parser.parse_args()

    try:
        if options.want_hist:
            cols = []
            max_val = 0
            col_names = []
            for i, f in enumerate(files):
                try:
                    (max_in_file, hist) = hist_file(f, 1)
                    cols.append([i + 1] + [to_str(x) for x in hist])
                    max_val = max(max_val, max_in_file)
                    col_names.append(f)
                except IOError, msg:
                    print >> sys.stderr, msg
            bins = ['Bin'] + make_bins(max_val)
            print_cols([bins] + cols)
            print '# Columns:'
            for i, f in enumerate(col_names):
                print '# (%d) %s' % (i + 1, f)
        elif options.want_percentiles:
            cols = []
            percentiles = make_percentiles(options.resolution)
            col_names = []
            for i, f in enumerate(files):
                try:
                    percs = percentiles_file(f, percentiles)
                    cols.append([i + 1] + [to_str(x) for x in percs])
                    col_names.append(f)
                except IOError, msg:
                    print >> sys.stderr, msg
            perc = ['Percentile'] + [to_str(x) for x in percentiles]
            print_cols([perc] + cols)
            print '# Columns:'
            for i, f in enumerate(col_names):
                print '# (%d) %s' % (i + 1, f)
        else:
            rows = []
            rows.append(STATS_HEADERS)
            for f in files:
                try:
                    rows.append(get_stats(f))
                except IOError, msg:
                    print >> sys.stderr, msg
            print_rows(rows)
    except KeyboardInterrupt:
        pass