#!/usr/bin/env python
+#---------------------------------------------------------------------------------------------------
+# Example invokation:
+# % ./regress.py griffon_skampi_pt2pt.ski.dat 65536 120832
+#
+#
+# Given two vectors of same length n: message size S(.. s_i ..), and communication time T( .. t_i .. )
+# where t_i is the time associated to a mesage size s_i, computes the segmentation of the vectors
+# in 3 segments such that linear regressions on the 3 segments maximize correlation.
+# The metric for correlation is now the cumulative, log error of the estimated time value given
+# by regression, to the real time value.
+#
+# Call r(X[i:j],Y[i:j]) = ax + b the regression line for data of X,Y between indices i and j
+# Call E[i:j] = E( e_i, .., e_j) the vector of estimates from the regression, where e_i = a*S[i] + b
+# Call mean_logerr( T,T' ) the average log error of paired elements t_i and t'_i, of vectors len n T and T' resp.
+# i.e mean_logerr( T,T' ) = 1/n * Sum_i^n ( exp(| ln(t_i)-ln(t'_i)|) - 1 )
+#
+# The script computes indices k and l, s.t.
+# mean_logerr( r(S[0:k],T[0:k]) , E[0,k] ) +
+# mean_logerr( r(S[k:l],T[k:l]) , E[k,l] ) +
+# mean_logerr( r(S[l:n],T[l:n]) , E[l,n] )
+# is minimum.
+#---------------------------------------------------------------------------------------------------
+
import sys
-from math import sqrt
+from math import sqrt,log,exp
if len(sys.argv) != 2 and len(sys.argv) != 4:
- print("Usage : %s datafile", sys.argv[0])
- print("or : %s datafile p1 p2", sys.argv[0])
+ print("Usage : {} datafile".format(sys.argv[0]))
+ print("or : {0} datafile p1 p2".format(sys.argv[0]))
print("where : p1 < p2 belongs to sizes in datafiles")
sys.exit(-1)
##-------------------------------------------------
## cov : covariance
## param X first data vector (..x_i..)
-## param Y second data vector (..x_i..)
+## param Y second data vector (..y_i..)
## = 1/n \Sum_{i=1}^n (x_i - avg(x)) * (y_i - avg(y))
##--------------------------------------------------
def cov(X,Y):
-
+ assert(len(X)==len(Y))
n=len(X) # n=len(X)=len(Y)
avg_X = avg( X )
avg_Y = avg( Y )
##----------------------------------
## variance : variance
## param X data vector ( ..x_i.. )
-## (S_X)^2 = (Sum ( x_i - avg(x) )^2 ) / n
+## (S_X)^2 = (Sum ( x_i - avg(X) )^2 ) / n
##----------------------------------
def variance( X ):
S_X2 = 0
return (S_X2/n)
+##----------------------------------
+## mean log_error
+## param X data vector ( ..x_i.. ), length n
+## param Y data vector ( ..y_i.. ), length n
+## return mean( 1/n * Sum_i^n ( exp(| ln(x_i)-ln(y_i)|) - 1 )
+##----------------------------------
+def mean_logerr( X,Y ):
+ assert( len(X) == len(Y) )
+ E = list(); # the list of errors
+ for i in range(len(X)):
+ E.append( exp(abs(log(X[i])-log(Y[i])))-1 )
+ return (avg( E ))
+
+
+##-----------------------------------------------------------------------------------------------
+## correl_split_weighted_logerr : compute regression on each segment and
+## return the weigthed sum of correlation coefficients
+## param X first data vector (..x_i..)
+## param Y second data vector (..x_i..)
+## param segments list of pairs (i,j) where i refers to the ith value in X, and jth value in X
+## return (C,[(i1,j1,X[i1],X[j1]), (i2,j2,X[i2],X[j2]), ....]
+## where i1,j1 is the first segment, c1 the correlation coef on this segment, n1 the number of values
+## i2,j2 is the second segment, c2 the correlation coef on this segment, n2 the number of values
+## ...
+## and C=c1/n1+c2/n2+...
+##-----------------------------------------------------------------------------------------------
+def correl_split_weighted_logerr( X , Y , segments ):
+ # expects segments = [(0,i1-1),(i1-1,i2-1),(i2,len-1)]
+ correl = list()
+ interv = list() # regr. line coeffs and range
+ glob_err=0
+ for (start,stop) in segments:
+ #if start==stop :
+ # return 0
+ S_XY= cov( X [start:stop+1], Y [start:stop+1] )
+ S_X2 = variance( X [start:stop+1] )
+ a = S_XY/S_X2 # regr line coeffs
+ b = avg ( Y[start:stop+1] ) - a * avg( X[start:stop+1] )
+ # fill a vector (Z) with predicted values from regression
+ Z=list()
+ for i in range(start,stop+1):
+ Z.append( a * X[i] + b )
+ # compare real values and computed values
+ e = mean_logerr( Y[start:stop+1] , Z )
+ correl.append( (e, stop-start+1) ); # store correl. coef + number of values (segment length)
+ interv.append( (a,b, X[start],X[stop],e) );
+
+ for (e,l) in correl:
+ glob_err = glob_err + (e*l/len( X )) # the average log err for this segment (e) is
+ # weighted by the number of values of the segment (l) out of the total number of values
+
+ #print("-> glob_corr={}\n".format(glob_corr))
+ return (glob_err,interv);
+
##-----------------------------------------------------------------------------------------------
## correl_split_weighted : compute regression on each segment and
## return the weigthed sum of correlation coefficients
c = S_XY/(sqrt(S_X2)*sqrt(S_Y2))
a = S_XY/S_X2 # regr line coeffs
b= avg ( Y[start:stop+1] ) - a * avg( X[start:stop+1] )
- print(" range [%d,%d] corr=%f, coeff det=%f [a=%f, b=%f]" % (X[start],X[stop],c,c**2,a, b))
+ #print(" range [%d,%d] corr=%f, coeff det=%f [a=%f, b=%f]" % (X[start],X[stop],c,c**2,a, b))
correl.append( (c, stop-start) ); # store correl. coef + number of values (segment length)
interv.append( (a,b, X[start],X[stop]) );
for (c,l) in correl:
glob_corr = glob_corr + (l/sum_nb_val)*c # weighted product of correlation
- print('-- %f * %f' % (c,l/sum_nb_val))
+ #print('-- %f * %f' % (c,l/sum_nb_val))
- print("-> glob_corr=%f\n" % glob_corr)
+ #print("-> glob_corr={}\n".format(glob_corr))
return (glob_corr,interv);
c = S_XY/(sqrt(S_X2)*sqrt(S_Y2))
a = S_XY/S_X2 # regr line coeffs
b= avg ( Y[start:stop+1] ) - a * avg( X[start:stop+1] )
- print(" range [%d,%d] corr=%f, coeff det=%f [a=%f, b=%f]" % (X[start],X[stop],c,c**2,a, b))
+ #print(" range [%d,%d] corr=%f, coeff det=%f [a=%f, b=%f]" % (X[start],X[stop],c,c**2,a, b))
correl.append( (c, stop-start) ); # store correl. coef + number of values (segment length)
interv.append( (a,b, X[start],X[stop]) );
for (c,l) in correl:
glob_corr = glob_corr * c # product of correlation coeffs
- print("-> glob_corr=%f\n" % glob_corr)
return (glob_corr,interv);
## ---------------
#count= 8388608 8388608 144916.1 7.6 32 144916.1 143262.0
#("%s %d %d %f %f %d %f %f\n" % (countlbl, count, countn, time, stddev, iter, mini, maxi)
- readdata.append( (int(l[1]),float(l[3]) / 2 ) ); # divide by 2 because of ping-pong measured
+ readdata.append( (int(l[1]),float(l[3])) );
nblines=nblines+1
## These may not be sorted so sort it by message size before processing.
sorteddata = sorted( readdata, key=lambda pair: pair[0])
sizes,timings = zip(*sorteddata);
+# zip makes tuples; cast to lists for backward compatibility with python2.X
+sizes = list(sizes)
+timings = list(timings)
##----------------------- search for best break points-----------------
## example
p1inx = sizes.index( p1 );
p2inx = sizes.index( p2 );
- max_glob_corr = 0;
+ max_glob_corr = 999990;
max_p1inx = p1inx
max_p2inx = p2inx
## tweak parameters here to extend/reduce search
- search_p1 = 30 # number of values to search +/- around p1
- search_p2 = 65 # number of values to search +/- around p2
+ search_p1 = 70 # number of values to search +/- around p1
+ search_p2 = 70 # number of values to search +/- around p2
min_seg_size = 3
+ if (search_p2 + min_seg_size > len(sizes)): # reduce min segment sizes when points close to data extrema
+ min_seg_size = len(sizes)-search_p2
+ if (search_p1 - min_seg_size < 0):
+ min_seg_size = search_p1
- lb1 = max(1, p1inx-search_p1)
- ub1 = min(p1inx+search_p1,search_p1, p2inx);
- lb2 = max(p1inx,p2inx-search_p2) # breakpoint +/- delta
- ub2 = min(p2inx+search_p2,len(sizes)-1);
+ lb1 = max( 1, p1inx-search_p1 )
+ ub1 = min( p1inx+search_p1, p2inx);
+ lb2 = max( p1inx, p2inx-search_p2) # breakpoint +/- delta
+ ub2 = min( p2inx+search_p2, len(sizes)-1);
print("** evaluating over \n");
print("interv1:\t %d <--- %d ---> %d" % (sizes[lb1],p1,sizes[ub1]))
print("rank: \t (%d)<---(%d)--->(%d)\n" % (lb1,p1inx,ub1))
print("interv2:\t\t %d <--- %d ---> %d" % (sizes[lb2],p2,sizes[ub2]))
print("rank: \t\t(%d)<---(%d)--->(%d)\n" % (lb2,p2inx,ub2))
+
+ result = list()
for i in range(lb1,ub1+1):
for j in range(lb2,ub2+1):
if i<j: # segments must not overlap
if i+1 >=min_seg_size and j-i+1 >= min_seg_size and len(sizes)-1-j >= min_seg_size : # not too small segments
- print("** i=%d,j=%d" % (i,j))
+ #print("** i=%d,j=%d" % (i,j))
segments = [(0,i),(i,j),(j,len(sizes)-1)]
- (glob_cor, interv) = correl_split( sizes, timings, segments)
- if ( glob_cor > max_glob_corr):
- max_glob_corr = glob_cor
- max_interv = interv
+ result.append( correl_split_weighted_logerr( sizes, timings, segments) ) # add pair (metric,interval)
+
+ # sort results on ascending metric: ok for logerr. Add "reverse=true" for desc sort if you use a correlation metric
+ result = sorted( result, key=lambda pair: pair[0])
+
+
+ top_n_sol=5; # tweak to display top best n solution
+
+ print("#-------------------- result summary ---------------------------------------------------------------------\n");
+
+ for k in range(top_n_sol):
+ (err,interval) = result[k]
+ print(k)
+ print("\n RANK {0}\n-------".format(k))
+ print("** overall metric = {0}".format(err))
+ for (a,b,i,j,e) in interval:
+ print("** OPT: [{0} .. {1}] segment_metric={2} slope: {3} x + {4}".format(i,j,e,a,b))
+
- for (a,b,i,j) in max_interv:
- print("** OPT: [%d .. %d]" % (i,j))
- print("** Product of correl coefs = %f" % (max_glob_corr))
+ print("\n\n\n")
- print("#-------------------- cut here the gnuplot code -----------------------------------------------------------\n");
+ print("#-------------------- Best Solution : cut here the gnuplot code -----------------------------------------------------------\n");
preamble='set output "regr.eps"\n\
set terminal postscript eps color\n\
set key left\n\
print(preamble);
print('plot "%s" u 3:4:($5) with errorbars title "skampi traces %s",\\' % (sys.argv[1],sys.argv[1]));
- for (a,b,i,j) in max_interv:
+ (err,interval) = result[0]
+ for (a,b,i,j,e) in interval:
print('"%s" u (%d<=$3 && $3<=%d? $3:0/0):(%f*($3)+%f) w linespoints title "regress. %s-%s bytes",\\' % (sys.argv[1],i,j,a,b,i,j))
print("#-------------------- /cut here the gnuplot code -----------------------------------------------------------\n");
S_X2 = variance( sizes )
S_Y2 = variance( timings ) # to compute correlation
- a = S_XY/S_X2;
+ a = S_XY/S_X2
correl = S_XY/(sqrt(S_X2)*sqrt(S_Y2)) # corealation coeff (Bravais-Pearson)
