CPSC 427: Object-Oriented Programming

Michael J. Fischer

Lecture 15
October 26, 2016

How to measure run time of a program

• There is no standard procedure in C++ for accurately measuring time.
• Time measurement depends on the software clocks provided by your computer and operating system.
• Clocks advance in discrete clicks called jiffies. A jiffy on the Zoo linux machines is one millisecond (0.001 seconds) long.
• Even if the clock is 100% accurate, the measured time can be off by as much as one jiffy.
• Hence, times shorter than tens of milliseconds cannot be directly measured with much accuracy using the standard software clock.

High resolution clocks

• Linux also provides high resolution clocks based on CPU timers.
• High resolution clocks are useful to the operating system for task scheduling and timeouts.
• They are also available to the user for higher-precision time measurements.
• Be aware that reading the clock involves a kernel call that takes a certain amount of time. This itself may limit the accuracy of timing measurements, even when the clock resolution is sufficiently high for the desired accuracy.
• See man 7 time for more information about linux clocks.

Measuring time in real systems

• Measuring code efficiency in real systems is challenging. Many factors can influence the results that are hard to control.
  • Other process running on the same machine.
  • Time spent in the OS moving data on and off disks.
  • Memory caching behavior.
• Lacking a controlled laboratory environment, one can still take steps to improve accuracy of tests:
  • Do some tests to determine what factors seem to have a sizable effect on the run time, e.g., the first run of a program is likely to be slower than subsequent runs because of caching.
  • Run the same test several times to get a feeling for the variance of results.
  • Make sure the optimizer isn’t optimizing away code that you think is being executed.

Realtime measurements StopWatch is a class I wrote for measuring realtime performance of code.

It emulates a stopwatch with 3 buttons: reset, start, and stop.

At any time, the watch displays the cumulative time that the stopwatch has been running.

HirezTime class

HirezTime is a wrapper class for the system-specific functions to read the clock.

It hides the details of the underlying time representation and provides a simple interface for reading, computing, and printing times and time intervals.

HirezTime objects are intended to be copied rather than pointed at, and they try to behave like other numeric types.

Versions of HirezTime

There are two versions: 15-StopWatch (Linux/Unix/MacOSX) Function gettimeofday() returns the clock in a struct timeval, which consists of two long ints representing seconds and microseconds. The resolution of the clock is system-dependent, typically 1 millisecond. (See demo 15-StopWatch.) 15-StopWatch-hirez (Linux only) Function clock_gettime() returns the clock in a struct timespec, which consists of two long ints representing seconds and nanoseconds. The resolution of the clock is system-dependent and can be obtained with the clock_getres() function. (See demo 15-StopWatch-hirez.)

HirezTime structure

• In C++, struct T and class T are very similar. In both cases, T becomes a new type name.
• struct members are public by default.
  class members are private by default.
• HirezTime is derived from struct timeval or struct timespec, depending on the version.
• It uses protected derivation to hide the underlying representation.
• It presents two interfaces to the world:

  1.

  The normal public interface treats HirezTime as an opaque object.

  2.

  A class derived from it can access the fields of the underlying timespec/timeval.

Printing a HirezTime number

Something seemingly simple like printing HirezTime values is not so simple. Naively, one might write:

where tv_sec and tv_usec are the seconds and microseconds fields of a timeval structure.

If t represents 2 seconds and 27 microseconds, then what would print is 2.27, not the correct 2.000027.

The class contains a print function that fixes this problem.

StopWatch class

StopWatch contains five member variables to remember

• Whether the watch is running or not.
• The cumulative run time to point when last stopped.
• The most recent start and stop times.

All functions are inline to minimize inaccuracies of measurement due to the overhead within the stopwatch code itself.

Casting a StopWatch to a HirezTime

An operator extension defines a cast for reading the cumulative time from a stop watch:

   operator HirezTime() const { return cumSpan; }

Thus, if sw is a StopWatch,

   cout << sw;

will print sw.cumSpan using sw.print().

Why it works

This works because operator<<() is not defined for righthand operands of type StopWatch but it is defined for HirezTime.

The compiler then coerces sw to something that is acceptable to the << operator.

Because operator HirezTime() is defined for class StopWatch, the compiler will invoke it to obtain a HirezTime object, for which << is defined.

Note that a similar coercion happens when one writes

   if(!in) {…}

to test if an istream object in is open for reading. Here, the istream object is coerced to a bool because operator bool() is defined inside the streams package.