1. Introduction

There are a variety of performance analysis tools available for examining the execution of a program:

• elapsed time: Measuring the CPU time and wall time of a program is the most important part of performance analysis. Unfortunately, this does not provide much insight about how performance can be improved. This can be measured on Unix systems with the time command.
• event counting: This is a generalization of elapsed time measurement. Instead of just counting clock ticks, however, on some architectures it is possible to count the number of floating point operations, the number of cache misses, etc. The PerfAPI project has designed a general API for all architectures that support performance monitoring.
• function call analysis: This usually involves recompiling the program with flags that cause code to be inserted around functions that keep track of the time spent in each function and the function call graph. This method is quite useful and available on many Unix systems. See the manual page for the gprof command for more information.
• statistical event sampling: This method causes periodic events to result in interrupts that record the current program counter. These program counters are then converted to the source code line numbers and the data is tabulated in a human readable form. Since this method would be too invasive if every event caused an interrupt, only every n'th event causes an interrupt and thus the method is not exact but statistical. Most programs spend a significant fraction of time in a few lines of code, so these statistical methods are usually accurate enough. Many Unix systems implement this functionality with the prof command. Most commonly, the events that are sampled are clock ticks. The PerfAPI project has developed a uniform interface to low level drivers that can statistically sample other events such as cache misses and flops on certain architectures. For Intel P6 and AMD K7 processors running Linux the perfctr package can be used to sample a wide range of events.

The visual profiler falls into the statistical event sampling category of profiling tools. It can sample clock ticks using the profil system call or can sample a wide range of system events using the PerfAPI package. Note that the PAPI software doesn't truly do statistical event sampling on all platforms. For these platforms, results for sampling events other than time will be inaccurate. If a Intel P6 or AMD K7 processor running Linux is used, then the perfctr package to accurately sample events other than clock ticks.