CS190
gprof assignment
April 17, 2001
Due Monday 4/30/01

The assignment

This assignment is intended to give you a chance to demonstrate that you know how to profile code and find performance bottlenecks.

Choose 5-10 methods from some component of your project. If you wrote code, use your code. Otherwise, pick any component you want, preferably aiming for one where performance will be important. Answer the following questions and record your answers before profiling.

1.

Before running gprof, make the following predictions:

• Rank the functions from most frequently called to least frequently called
• Rank the functions in terms of time spent in that function (``self'' time instead of cumulative time)

2.

Predict where the hot spots will be in the code you've chosen, and explain why.

3.

Compile your code without any optimizations and profile it with gprof. Save the output from gprof and answer the questions below.

4.

EXTRA CREDIT: Compile your code with optimizations (either -O2 or -fast will work, but check the compiler man page for the differences) and then profile it again. Do the same tasks so that the results are comparable. Save the output from gprof and answer the last question below

5.

Analyze the output from both runs, and answer the following questions

• What did you do with the application when you ran your program with profiling? Your choice of activities will likely affect what results you get, so it's important to perform a representative set of tasks.
• Where are the hot spots in your code? Were your predictions correct? Include hot spots in routines that you worked with as well as those that you did not.
• Which functions would be good candidates for optimization changes? Why?
• EXTRA CREDIT QUESTION: Suppose someone made the statement ``Having the compiler optimize your code simply makes everything run faster. All functions get called the same number of times, and take the same amount of time to execute relative to one another''. Is this statement accurate? What evidence (from the gprof output) do you have to support this? What evidence do you have to contradict this?

Record the answers to the above questions (both pre- and post-gprof), and include a subset of your gprof output. Include the sections of the call graph profile and the lines from the flat profile which contain the functions mentioned in both your predictions and your analysis. If you did the optimized vs. unoptimized run, include excerpts from both runs, labeled appropriately. Specify which level of optimization you used.

Due

Due Monday, April 30.

Using gprof

Here is a brief summary of how to use gprof.

• Recompile all your code for profiling. This means you should specify the -xpg option on the compile and link lines.
• Run the program. Perform a carefully crafted set of tasks that will exercise the parts of the program you want to evaluate. Having this be repeatable will make it much easier to evaluate any changes that you make to optimize the code. It must terminate normally (either by returning from main or calling exit(3c))
• run gprof -C [executable] to extract the profile.

The O'Reilly book High Performance Computing by Kevin Dowd and Charles Severance has a section on interpreting the output from gprof. Or see a TA!