1 #!/usr/bin/env python 2 3 import sys 4 from scipy.stats import mannwhitneyu 5 6 SIGNIFICANCE_THRESHOLD = 0.0001 7 8 a,b = {},{} 9 for (path, d) in [(sys.argv[1], a), (sys.argv[2], b)]: 10 for line in open(path): 11 try: 12 tokens = line.split() 13 samples = tokens[:-1] 14 label = tokens[-1] 15 d[label] = map(float, samples) 16 except: 17 pass 18 19 common = set(a.keys()).intersection(b.keys()) 20 21 ps = [] 22 for key in common: 23 _, p = mannwhitneyu(a[key], b[key]) # Non-parametric t-test. Doesn't assume normal dist. 24 am, bm = min(a[key]), min(b[key]) 25 ps.append((bm/am, p, key, am, bm)) 26 ps.sort(reverse=True) 27 28 def humanize(ns): 29 for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]: 30 if ns > threshold: 31 return "%.3g%s" % (ns/threshold, suffix) 32 33 maxlen = max(map(len, common)) 34 35 # We print only signficant changes in benchmark timing distribution. 36 bonferroni = SIGNIFICANCE_THRESHOLD / len(ps) # Adjust for the fact we've run multiple tests. 37 for ratio, p, key, am, bm in ps: 38 if p < bonferroni: 39 str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio 40 print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio) 41
