Concepts for Performance Testing
Before you
use performance tools to detect and correct performance bottlenecks in
your application, you should familiarize yourself with the various
terms used to describe data collected during profiling. The performance data gathered depends on the profiling method you choose - sampling or instrumentation.
Sampling
Sampling, in which the application is periodically interrupted,
provides the advantage of low overhead, which allows the application to
behave closer to how it would behave in the real world. During
sampling, the performance data collection infrastructure periodically
interrupts the application as it runs to determine what function is
running and increments the sample count of that function. It stores
information about the call stack leading up to the function call. The
drawback to this approach is that it can get only relative performance
data for the functions that were sampled. It is possible that a
function you wanted to sample did not get sampled and therefore, no
information is available about it.
Important
You must choose appropriate sampling event to be used during
sampling. For example, CPU cycle sample event only shows the locations
in the application that consume CPU cycles. If the application is
blocked waiting for disk, network, page fault, etc. the resulting
information will not be helpful in determining the actual problem. To
detect problems with page faults, use page fault as the sample event.
Instrumentation
Instrumentation provides the advantage of gathering exact
performance data for specific sections of the application. During
instrumentation, enter and exit probes are inserted into the
application's functions. These probes report back to the data
collection infrastructure and allow users to capture exact amounts of
time and other metrics that a function took to run.
Important
Probes are not inserted in inline functions - functions compiled
with /Ox (Full Optimization). Therefore, number of calls in the report
will not agree with the actual number of times the code block executed.
To determine the exact number of times that code block executed,
compile code with /Ob0. However, compiling code using this option will
impact code optimization.
After you profile your application, a report is generated. The
performance report file contains the data collected during profiling.
The following list provides terms that you will need to understand
before analyzing the report:
Application Time
Application time shows the time spent in the direct execution of
the profiled code. It excludes performance data that contain calls to
the operating system and time that was spent waiting for other threads
to execute ().
Elapsed Time
Elapsed time shows the total system time spent executing the
profiled code. It includes performance data that contain transition
events.
Exclusive
The term exclusive refers to only those samples taken in the
function, and does not include samples taken in other functions called
by it.
Inclusive
The term inclusive refers to the samples taken in the function, and includes the samples taken in other functions called by it.
Transition
The number of transition events that occurred during profiling the application.
Transition Events
A change in the location of processor event execution between ring
3 (user mode) and ring 0 (kernel-mode). Transition events represent
time spent outside the direct execution of the application code.
Transition events can be time spent in threads that are not part of the
profiled item, or time spent executing calls from the profiled item to
the operating system.
Allocations
Memory and type instances allocated during profiling the
application. The two type of allocation reported are: exclusive and
inclusive.
Bytes Allocated
Bytes allocated during profiling the application. The two type of byte allocation reported are: exclusive and inclusive.
Root
The term root refers to a function that calls and/or is called by
one or more functions. Information about root functions appears in the
Caller/Callee report and view. In the http://msdn2.microsoft.com/en-us/library/ms242754.aspx - Caller/Callee View , root function is listed in the middle part of the view.
|