How to average time commands

benchmarkperformancetimetime-utility

I am timeing some of my commands. Now to average the measures, I'd like to run and time my commands a specified number of times, and get the results with a calculated mean and standard deviation. The result would be like:

avgtime -n 100 my_command

real    0m5.388s stdev 0m0.068s
user    0m5.380s stdev 0m0.067s
sys     0m0.004s stdev 0m0.000s

Is there a UNIX tool for this?
Does GNU/Linux have one?

Best Answer

You can to try use the timeit module, available in any system with Python:

$ python -m timeit "__import__('os').system('my command here')"
10 loops, best of 3: 591 msec per loop

Related Solutions

Time – How to Get the Time of Several Commands with a Loop

You can write a simple for-loop

 time -p bash -c "for (( i=0; i<10; i++ )); do command1; command2; done;"

Note that I used bash instead of sh for the loop.

Python vs bc in Evaluating 6^6^6

Python imports a large number of files at startup:

% python -c 'import sys; print len(sys.modules)'
39

Each of these requires an even greater number of attempts at opening a Python file, because there are many ways to define a module:

% python -vv -c 'pass'
# installing zipimport hook
import zipimport # builtin
# installed zipimport hook
# trying site.so
# trying sitemodule.so
# trying site.py
# trying site.pyc
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site.so
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/sitemodule.so
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site.py
# /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site.pyc matches /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site.py
import site # precompiled from /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site.pyc
# trying os.so
# trying osmodule.so
# trying os.py
# trying os.pyc
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/os.so
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/osmodule.so
# trying /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/os.py
# /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/os.pyc matches /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/os.py
import os # precompiled from /System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/os.pyc
    ...

Each "trying", except for those which are builtin, requires an os-level/system calls, and each "import" seems to trigger about 8 "trying" messages. (There was ways to reduce this using zipimport, and each path in your PYTHONPATH may require another call.)

This means there are almost 200 stat system calls before Python starts on my machine, and "time" assigns that to "sys" rather than "user", because the user program is waiting on the system to do things.

By comparison, and like terdon said, "bc" doesn't have that high of a startup cost. Looking at the dtruss output (I have a Mac; "strace" for a Linux-based OS), I see that bc doesn't make any open() or stat() system calls of its own, except for loading a few shared libraries are the start, which of course Python does as well. In addition, Python has more files to read, before it's ready to process anything.

Waiting for disk is slow.

You can get a sense for Python's startup cost by doing:

time python -c pass

It's 0.032s on my machine, while 'print 6**6**6' is 0.072s, so startup cost is 1/2rd of the overall time and the calculation + conversion to decimal is the other half. While:

time echo 1 | bc

takes 0.005s, and "6^6^6" takes 0.184s so bc's exponentiation is over 4x slower than Python's even though it's 7x faster to get started.

Best Answer

Related Solutions

Time – How to Get the Time of Several Commands with a Loop

Python vs bc in Evaluating 6^6^6

Related Question