avgtime

Works like the linux time command, except it accepts a -r argument to specify repetitions and shows stats.

If repetitions are specified, then stats are computed and shown, like median, average, and standard deviation.

Pass -q to make the program under test silent (stdout and stderr piped to /dev/null).

Pass -d to discard the first run. If -d is passed, there is one extra repetition which is discarded.

Pass -p to print the sorted times in milliseconds.

Pass -h to print a very nice histogram.

How to use

Do:

    avgtime -r 1000 <your_command>

to run <your_command> a thousand times.

For instance:

    /avgtime -q -r30  ls -lR /etc

    ------------------------
    Total time (ms): 2914.28
    Repetitions    : 30
    Sample mode    : 99 (6 ocurrences)
    Median time    : 98.01
    Avg time       : 97.1426
    Std dev.       : 4.52638
    Minimum        : 89.625
    Maximum        : 106.68
    95% conf.int.  : [88.2711, 106.014]  e = 8.87154
    99% conf.int.  : [85.4834, 108.802]  e = 11.6592
    EstimatedAvg95%: [95.5229, 98.7623]  e = 1.61972
    EstimatedAvg99%: [95.0139, 99.2712]  e = 2.12867

Displayed times are in milliseconds.

Run avgtime without arguments to see more usage help.

Don't forget to checkout the nice histogram.

Note that the following histogram shows two peaks because of the CPU speed switching that laptops do. Some laptops will often switch to low speed even in "High Performance" mode and plugged to the wall. Forcing to high speed (always on) eliminates the second peak (how to do that is beyond the scope).

    avgtime -q -h -r1000   ls /etc -lR

    ------------------------
    Total time (ms): 97920.7
    Repetitions    : 1000
    Sample mode    : 90 (137 ocurrences)
    Median time    : 99.0305
    Avg time       : 97.9207
    Std dev.       : 5.26182
    Minimum        : 89.081
    Maximum        : 111.864
    95% conf.int.  : [87.6077, 108.234]  e = 10.313
    99% conf.int.  : [84.3672, 111.474]  e = 13.5535
    EstimatedAvg95%: [97.5945, 98.2468]  e = 0.326125
    EstimatedAvg99%: [97.4921, 98.3493]  e = 0.428601
    Histogram      :
        msecs: count  normalized bar
           89:    45  #############
           90:   137  ########################################
           91:    58  ################
           92:    43  ############
           93:    16  ####
           94:     8  ##
           95:    33  #########
           96:    44  ############
           97:    59  #################
           98:    53  ###############
           99:    76  ######################
          100:    84  ########################
          101:    87  #########################
          102:    86  #########################
          103:    77  ######################
          104:    35  ##########
          105:    22  ######
          106:    12  ###
          107:    10  ##
          108:     6  #
          109:     6  #
          110:     1  
          111:     2  

You can see the importance of the histogram now, and of statistics such as the 'sample mode', 'median', and standard deviation of the sample. All these together might help visualize and quantify performance effects that you wouldn't see by timing your program just once.

How to install

This little program is written in the most awesome programming language of the world: D (http://dlang.org).

To compile avgtime, you must have D installed. You can download D from http://dlang.org/download.html

Compile with:

    dmd avgtime.d

and that's it. You can use ./avgtime now. Copy it to /usr/local/bin to have it in your path.

Why do I need the stats?

If you are a benchmarker type of guy, you know that there is no fast or slow. There is only faster and slower. So you usually find yourself comparing the times of two different things, to see which one is faster. Do you run them once? Are you satisfied that you improved the running time of your program with just one run?

That depends on how big the time difference is. And usually, we optimize things one step at a time, making small improvements. This all means that we must run the program many times to distinguish our little improvements from background noise. Which now means that the running time of your program has become a random variable (in the probability theory sense), of which we must take samples to reach a conclusion about its probability distribution.

Even when doing dramatic optimizations, as opposed to little ones, ie: reducing time complexity by using better algorithms, we must make sure that it was worth it and keep constant times in check.

To really know whether you made an improvement on running time, you must take into account the standard deviation (or any other confidence interval), and at least make sure that the Average +/- StdDev intervals don't overlap. Or, if you are too lazy, just make sure that the maximum time of the 'fast' version, is better than the minimum of the 'slow' version under test.

See, one is not measuring the running time of a program. One is actually measuring the interval of running times where it is likely to always stay... to know more about its probability distribution.

(Look for 'Hypothesis Testing' and 'Confidence Intervals' on the web for more on the subject).

HAVE FUN!

--