CodeXL User GuideCodeXL handles all of the details and mechanics of data collection and profile formation.
CodeXL uses statistical sampling to collect and build a program profile. Because all profiles rely on statistical sampling, it is important for the CPU Profiler to take enough samples. The number of samples collected during a session depends on the following:
· how frequently samples are taken (the sampling period or interval)
· the measurement length; that is, the amount of time which takes for the CPU Profiler to take a specific sample
· whether the profiling is system-wide
CodeXL provides the following profile sessions types:
Investigate Instruction Access
The type and frequency of the profiled events can indicate the presence of a pipeline bottleneck, poor memory access pattern, poorly predicted conditional branches, or some other performance issues. Once hot-spots are found through time-based profiling, EBP and IBS are used identify opportunities for optimization. AMD processors provide a wide range of hardware events that can be monitored and measured. The number of counters and the hardware events that can be measured are processor-dependent.
When compiling your application to use for profiling, consider generating debug information as well. If present, the debug information can help identify profiled areas in the code; however, debug information is not required. Performance data can be collected for an application program that was compiled without debug information; in this case the results displayed by CodeXL are less descriptive. For example, CodeXL displays assembly code rather than function names or source code.
The CPU Profiler measures CPU execution time of a program. Through these measurements it helps you optimize the program. The Profiler records which functions were opened and how long it took to execute each one. CodeXL does this by taking samples. To take a sample, CodeXL interrupts the program’s execution at specified intervals, and logs the state of the program’s call chain.
CodeXL samples all of the various performance monitoring registers to obtain detailed information about the running application or the entire system. While all work can be done through the GUI, profiles also can be collected and analyzed through Visual Studio by using the CodeXL Visual Studio extension.
System profiling identifies a hot-spot anywhere in the system or an application under test. Any software component (an executable image, dynamically loaded library, device driver, or even the operating system kernel) that executes during the measurement period can be sampled. Any child processes spawned by a profiled process are profiled automatically. During the profiling process, the application to be analyzed is run at full speed on the same machine that is running CodeXL.
Time-based samples (collected at 1 ms intervals on each core) can be used to identify possible bottlenecks, execution penalties, or optimization opportunities. The TBP feature can be used on both AMD and non-AMD processors with an Advanced Programmable Interrupt Controller (APIC) timer. Event-based samples and instruction-based samples can be used to help determine the cause of hot-spots or optimization opportunities. The sampling intervals are weighted so the types of view analysis are valid. The EBP and IBS features are only available on AMD processors.
Call chain sampling collects function call information, including caller-to-called relationships between functions. It is used in conjunction with the selected profile. When call chain sampling is enabled, CodeXL collects information from the run-time call stack of a monitored application process (and child processes) whenever a regular profile sample is taken for the process. When compared to other techniques, such as instrumentation, call chain sampling is a relatively low-overhead approach to the collection of function call information. However, call chain sampling results are subject to statistical variation.