gprof, bprof and Time Profilers

A friend of mine was explaining to me why he thought his program wasn't running fast enough. “Have you profiled it?” I asked. “No, but I'm pretty sure I know where the bottleneck is,” he replied—famous last words. “Well, let's try the profiler,” I said. Profiling quickly revealed that 98% of the CPU was being spent in one subroutine of my friend's program and that 86% of the CPU was being spent in one line, and he was wrong about the location of the bottleneck.

Profilers are invaluable tools that let you know where a program is spending most of its time. This information is extremely valuable because it tells you where your time can best be spent in making your program more efficient and where you are wasting your time. Typical programs are not quite as lopsided as in the above story. The “80-20” rule says that a program will spend 80% of its time in about 20% of the code.

If the total running time of a program is n, then we can break up the running time into pieces:

total-time = a1*n + a2*n + a3*n + ...

where the “a”s represent fractions of the total time that your program spends in a particular segment. The sum of the “a”s must add to one. The 80-20 rule says that one of the “a”s will be quite large. For example, suppose a1 is 8/10 and a2 is 2/10. These numbers correspond to a program which spends 80% of its time doing a1 and 20% for everything else.

total-time = .8*n + .2*n

Now suppose we optimize a2 so that it runs twice as fast as before—a significant speedup. The time 0.2*n is now 0.1*n. The total running time is now 0.8n+0.1n = 0.9n, meaning the whole program executes in 90% of the time that it originally did. Suppose we instead concentrate on the other piece. If we halve the running time of the first piece, it becomes 0.4n. 0.4n+0.2n=0.6n, or 60% of the original running time. As you can see, it is worth our while to concentrate on the particular portion of the program that dominates the runtime.

In my friend's case, we were able to optimize that single line a bit. The real optimization came two days later when he told me that he had removed the whole subroutine in question, simply by changing how he thought about his data.

The easiest profiler to use under Linux is the gprof profiler. gprof is a standard part of the GNU development tools. If you have gcc installed, you probably have gprof too. To use gprof, simply recompile your program with gcc using the -pg switch. This option causes gcc to insert a bit of extra code into the beginning of each subroutine in your program. The -pg switch must also be used when you link your program, since another snippet of code must be present to tie the pieces together.

After recompiling, run your program. It will execute slightly slower because of the work needed to profile the code, but it shouldn't be too slow. After the program finishes, there will be a file named gmon.out in the current directory. This file contains the profiling information collected during the program's run. gprof is used to print this in a human readable form.

gprof outputs information in two ways: a flat profile and a call graph. The flat profile tells you how much time the program spent in all of the subroutines, and the call graph tells you which subroutines called which subroutines.

The first part of a flat profile is shown in Listing 1. The “self seconds” column shows how much time was taken up by each subroutine. The number of times each subroutine was invoked is shown in the “calls” column. The “self ms/call” columns gives the average time in milliseconds spent in a given subroutine, while the “total ms/call” includes time spent in subroutines called by that subroutine as well. For reasons explained in the gprof documentation, this last column is actually only a guess and should not be relied upon.

In this example, an important thing to note is that the mcount subroutine is the actual profiling subroutine call inserted by gcc when it compiles code with the -pg switch. The fact that the program spends nearly 20% of its time here indicates that a lot of calling and returning is happening, and that one way to speed up the code would be to eliminate some of the subroutine calls. Which ones? The subroutines rnd and uni are likely candidates, since they are called 142 million times in 900 seconds.

The other profiler in common use on Linux is bprof. The major difference between bprof and gprof is that bprof gives timings on a source line basis while gprof has only subroutine-level resolution, and also includes information like invocation counts. To use bprof, link an object file, bprof.o, into your program. After you've run your program, a file named bmon.out contains the timing information. Run bprof on this data file, and it makes copies of your source files with timing numbers prepended to each line.