Weird stat numbers
rzezeski opened this issue · 0 comments
rzezeski commented
While playing around with batch sizes I noticed odd results:
[root] # bin/mprotect -E -C 200 -B 1 -L -S -W -N mprot_twz8k -l 8k -I 1000 -w -t -f /dev/zero
Running: mprot_twz8k for 0.01735 seconds
# bin/../bin-i386/mprotect -E -C 200 -B 1 -L -S -W -N mprot_twz8k -l 8k -I 1000 -w -t -f /dev/zero
prc thr usecs/call samples cnt/samp size flags
mprot_twz8k 1 1 7.67600 1486 1 8192 --w-t
#
# STATISTICS usecs/call (raw) usecs/call (outliers removed)
# min 0.73000 0.73000
# max 41.49400 14.10400
# mean 4.31983 4.29481
# median 7.67600 7.67600
# stddev 3.67108 3.54326
# standard error 0.09520 0.09192
# 99% confidence level 0.22144 0.21380
# skew 0.71464 0.03937
# kurtosis 5.03037 -1.89294
# time correlation -0.00009 0.00001
#
# elasped time 0.01079
# number of samples 1486
# number of outliers 1
# getnsecs overhead 624
#
# DISTRIBUTION
# counts usecs/call means
# 734 0.00000 |******************************** 0.75871
# 5 1.00000 |* 1.33680
# 4 2.00000 |* 2.21900
# 0 3.00000 | -
# 0 4.00000 | -
# 0 5.00000 | -
# 0 6.00000 | -
# 668 7.00000 |***************************** 7.75380
#
# 75 > 95% |*** 8.40128
#
# mean of 95% 4.07654
# 95th %ile 7.82300
[root] # bin/mprotect -E -C 200 -B 1 -L -S -W -N mprot_twz8k -l 8k -I 1000 -w -t -f /dev/zero
Running: mprot_twz8k for 0.01743 seconds
# bin/../bin-i386/mprotect -E -C 200 -B 1 -L -S -W -N mprot_twz8k -l 8k -I 1000 -w -t -f /dev/zero
prc thr usecs/call samples cnt/samp size flags
mprot_twz8k 1 1 0.87200 1282 1 8192 --w-t
#
# STATISTICS usecs/call (raw) usecs/call (outliers removed)
# min 0.72800 0.72800
# max 1020.97300 10.01900
# mean 5.16415 4.24617
# median 7.69200 0.87200
# stddev 28.56642 3.51713
# standard error 0.79505 0.09823
# 99% confidence level 1.84928 0.22848
# skew 34.83146 0.01390
# kurtosis 1235.51234 -1.99846
# time correlation 0.00187 -0.00000
#
# elasped time 0.01083
# number of samples 1282
# number of outliers 9
# getnsecs overhead 625
#
# DISTRIBUTION
# counts usecs/call means
# 643 0.00000 |******************************** 0.75162
# 2 1.00000 |* 1.23950
# 0 2.00000 | -
# 0 3.00000 | -
# 0 4.00000 | -
# 0 5.00000 | -
# 0 6.00000 | -
# 572 7.00000 |**************************** 7.76783
#
# 65 > 95% |*** 7.91723
#
# mean of 95% 4.05010
# 95th %ile 7.81500
- While the distribution look pretty close the
usecs/call
are an order of magnitude different. Perhaps using this stat inmultiview
is a bad idea considering how much it can vary from run to run. - Some of the
usecs/call
stats are negative. - Perhaps distribution should be log-lin histogram like DTrace?