August 29, 2007
Introduction to Power DebuggingPoor Man's Performance Timing
There have been times that I've written some code that I would have liked to get performance timing on. There have even been times when I wanted to get performance of an application that I didn't have the source for. The application may have been crunching numbers or talking with hardware and through debugging I had determined the locations I needed to measure. I just needed a way that I could do the measuring.
It is easy when you have the source to just modify the code with QueryPerformanceCounter and OutputDebugString and then recompile. After all, I just want an estimate and the code I'm attempting to measure isn't a few instructions but rather a more complex operation. The performance will always vary depending on the machine and the system load so to the nanosecond accuracy isn't necessary.
The flaw with modifying the code is that it would only work for that piece of code. It wouldn't work with applications I don't have the source to and I would be required to modify all locations that I want to time. I could buy a tool, but perhaps the tool is overkill for what I need and I don't want to spend the money. What about writing a debug extension? The real question is how much accuracy is lost when using the debug extension as opposed to modifying the source code. In order to find out, we need to measure the timing of some simple code, shown in Example 1.
for(uiIndex = 0;uiIndex < g_Loops; uiIndex++)
{
dSampleX *= dSampleY;
}
Example 1: Double precision multiplication in a loop.
The data shown in Table 1 is a comparison of two different methods to retrieve performance data. The first listed is "QPC" which is short for "QueryPerformanceCounter" API. This is an abstraction over an operating system-implemented method for querying a performance counter. The second is RDTSC and while the first method could possibly use RDTSC as well, there would be some more overhead involved over using the instruction yourself. The first method, however, could be implemented in a variety of methods including ACPI hardware ports.
Table 1: Timing comparisons
Accuracy and precision are the two factors that make up performance timing. Accuracy is how close your numbers are to the actual time and precision is the interval in which you are measuring. In the above, our precision is milliseconds and when doing performance monitoring via a debug extension I would only suggest to measure items that would require no less than a millisecond resolution.
Table 2: Performance-monitoring overhead
The data listed in Table 2 is the performance timing of doing nothing and essentially illustrates how much overhead each method typically has. Example output is shown in Figure 3 when using the RDTSC debug extension.
[Click image to view at full size]
Figure 3: Using the RDTSC debug extension.
The following are some of the key factors when interpreting the performance numbers:
In my timing you will note that I generally did not get one straight time for any of these samples. In fact they can vary by several milliseconds due to context switching, for example. You will notice that as the operation you are timing becomes longer the less effect the overhead of using a debug extension skews the results. In fact you may notice that in some cases the result was actually lower with the debug extension than the modified code! This may be a result of the debug extension running in the context of another thread and as such not sharing its time slice with the thread being timed.
This example demonstrates that if you are looking simply for estimates within a few milliseconds then a debug extension may just do the trick. However, if you need more accurate timing then a more isolated environment using increased thread priorities and possibly specialized tools may be a more appropriate solution.
Previous Page |
1 Power Debugging
|
2 Debugger Feature: Break Point Command Strings
|
3 Automatic Break Points
|
4 Debugger Feature: Custom Debug Extensions
|
5 Poor Man's Performance Timing
|
6 Debugger Feature: Symbol Search
Next Page
|
 |
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|
