Difference between revisions of "Linux Tools Project/Valgrind/User Guide"
m (→Exporting Valgrind Data)
(→Analyzing Cachegrind Profile Results)
|Line 148:||Line 148:|
== Analyzing Cachegrind Profile Results ==
== Analyzing Cachegrind Profile Results ==
The results of a Cachegrind profile run are displayed in the Valgrind view. These results show Cachegrind's cache/branch data in different levels of granularity. Double-clicking on any file, function, or line will open the corresponding source file and place the
The results of a Cachegrind profile run are displayed in the Valgrind view. These results show Cachegrind's cache/branch data in different levels of granularity. Double-clicking on any file, function, or line will open the corresponding source file and place the on the appropriate (if can be resolved).
Revision as of 11:11, 29 July 2009
- 1 Overview
- 2 Installing
- 3 General Usage
- 4 Using Memcheck
- 5 Using Massif
- 6 Using Cachegrind
- 7 Special Cases
Valgrind is an instrumentation framework for building dynamic analysis tools that can be used to profile applications in detail. Valgrind tools are generally used to automatically detect many memory management and threading problems. The Valgrind suite also includes tools that allow you to build new profiling tools to suit your needs.
The Valgrind plug-in for Eclipse (documented herein) integrates several Valgrind tools into Eclipse. This allows Eclipse users to seamlessly include profiling capabilities into their workflow. At present, the Valgrind plug-in for Eclipse supports three Valgrind tools: Memcheck, Massif, and Cachegrind.
For more information about Valgrind, refer to http://www.valgrind.org/.
In order for the Valgrind plug-in for Eclipse to work properly, you should have the valgrind-3.3.0 (or later) package installed on your system first.
Once the valgrind package is installed, the easiest way to install the Valgrind plug-in for Eclipse is through the Software Updates and Add-ons menu. For information on how to use this menu, refer to this link.
You can use the Valgrind plug-in's "one-click launch" shortcut to run a default profile. Doing so is similar to running
valgrind ./program_name from the command line. To use the one-click launch shortcut, right-click on the project or binary name in the Project Explorer window and navigate to Profile As > Profile With Valgrind. If a source file belonging to that program is available on any open editor, you can also right-click inside the editor to navigate to the one-click launch.
Configuring a Profile Run
The Profile Configurations menu allows you to configure a profile run with some basic Valgrind profiling options, along with the most useful options for each Valgrind tool. These settings can be configured in the Valgrind Options tab of a specific profile run.
The General Options tab allows you to configure the following options:
- Tool to run allows you to choose between Memcheck, Massif, and Cachegrind in your profile run. By default, using the one-click shortcut will run the Memcheck tool.
- Trace children on exec configures whether or not to trace into sub-processes initiated via the
execsystem call. This is identical to using the
--trace-children=option. This is disabled by default.
For more details on the use of the Trace children on exec option for each Valgrind tool, refer to Profiling Child Processes.
- Run __libc_freeres on exit configures whether or not to run a
__libc_freeresroutine after the profile run exits. This routine instructs
glibcto release all memory it has allocated, and is similar to using the
--run-libc-freeres=option. This is enabled by default.
- Child silent after fork is enabled by default, and in most cases cannot be configured. This option disables any debugging or logging output from child processes created from a
- Demangle C++ names allows Valgrind to translate encoded C++ names ("mangled" by
g++during compilation) back to something similar to their original form. This is enabled by default.
- Limit errors reported instructs Valgrind to stop reporting errors after a total of 10,000,000 actual errors or 1,000 unique errors have been detected. This is enabled by default.
- Callers in stack trace configures how many levels of function call names Valgrind should use to identify program locations. By default, Valgrind uses 12 levels.
- Show errors below main configures stack traces for errors should show any functions that appear beneath
main(). This is disabled by default.
- Max stack frame size is the maximum size of a stack frame, and is set to 2000000 by default.
For more information about these options, refer to
The Workspace button allows you to select a resource from the workspace as your suppressions file. To use a suppressions file outisde of the workspace, use the File System button.
Exporting Valgrind Data
To export the raw data collected during a Valgrind profile run, use the Export wizard menu. To access the Export wizard menu, navigate to File > Export and select Valgrind Log Files under the category Other.
Memcheck is the default tool used by Valgrind and the Valgrind plug-in for Eclipse. It allows you to detect memory-related problems such as memory leaks, use of uninitialized memory, and reads from/writes to inappropriate stack areas. For more information about Memcheck, refer to http://www.valgrind.org/docs/manual/mc-manual.html.
Analyzing Memcheck Profile Results
Any memory management errors detected by Memcheck are displayed in the Valgrind view, which appears automatically after a profile run. Each displayed error can be expanded in the Valgrind view to display its stack trace.
In some cases, Memcheck can determine which section of the source code causes a reported error. Whenever this is the case, double-clicking on the stack trace of the reported error will open the related source file and place the insertion point on the line responsible for the error.
Configuring a Memcheck Profile Run
The Memcheck Options tab allows you to configure the following Memcheck options:
- leak resolution sets how tolerant Memcheck should be in considering different stack traces to be the same. The default setting is "low", which means only the first two entries need to match. The "med" setting requires four entries to declare a match. The "high" setting requires all entries to declare match.
- freelist size (blocks) is identical to the Memcheck command-line option
--freelist-vol=, which specifies the maximum total size (in bytes) of memory blocks to be marked "inaccessible" once they have been freed using
free(as in C) or
delete(as in C++). The default value for freelist size (blocks) is 10000000.
- minimum heap block alignment is identical to the Memcheck command-line option
--alignment=, which specifies the minimum alignment/size of heap blocks. The default value for this option is 8.
- show reachable blocks configures whether or not Memcheck should report blocks it could not find a pointer to during leak detection. This is disabled by default (i.e. do not report blocks with undetermined pointers).
- undef value errors configures whether or not Memcheck should detect dangerous uses of undefined value errors. This is enabled by default.
- allow partial loads is identical to the Memcheck command-line option
--partial-loads-ok=, which controls how Memcheck handles word-sized, word-aligned loads from addresses from which bytes are addressable and others are not. This is disabled by default (i.e. treat loads from partially invalid and completely invalid addresses as the same).
- gcc 2.96 workarounds is identical to the Memcheck command-line option
--workaround-gcc296-bugs, which controls whether Memcheck should assume that small reads/writes below the stack pointer are due to bugs in gcc-2.96. This option is disabled by default (i.e. do not assume them to be reslting from gcc-2.96 bugs; report such reads/writes as errors).
Refer to http://www.valgrind.org/docs/manual/mc-manual.html#mc-manual.suppfiles for information about Memcheck suppression files. For more information about each Memcheck option in the Valgrind plug-in for Eclipse, refer to
Massif is a Valgrind tool that measures how much heap memory an application uses. Heap memory profiling is useful in determining how to reduce the memory usage of an application. For more information about Massif, refer to http://www.valgrind.org/docs/manual/ms-manual.html.
Analyzing Massif Profile Results
Massif distills a large volume of data. The Valgrind plug-in for Eclipse presents three ways to view this data:
- Snapshots Table. Each entry in the Valgrind view contains useful information about each snapshot taken by Massif during the profile run. The peak snapshot is highlighed in bold text.
- Heap Tree. Detailed snapshots are denoted by the Show Heap Tree icon ( ). Double-clicking any of these snapshots will modify the Valgrind view to display all detailed snapshots. Collapsing each detailed snapshot in this view will display a heirarchy of function calls (i.e. heap tree) detailing the heap allocations used during the snapshot.
Some functions within the heap tree list a related source file. Double-clicking these functions will open the listed source file and place the insertion point on the specific function responsible for the call.
You also can also toggle the Show Heap Tree shortcut on the Valgrind view toolbar to switch between Snapshots Table and Heap Tree views.
- Heap Chart. The information in the snapshots table is also available in line chart form, displayed in the Heap Chart window. Clicking any data point in the chart will highlight its corresponding snapshot in the snapshot table. Double-clicking on a data point that corresponds to a detailed snapshot will open an editor to one of its function calls.
If a detailed snapshot contains calls to multiple source files (or multiple functions within a source file), double-clicking its corresponding data point on the heap chart will open the Open Allocation Function Call menu. Double-clicking a function from the Open Allocation Function Call menu will open its corresponding source file and place the insertion point on the specific function responsible for the call.
Configuring a Massif Profile Run
The Massif Options tab allows you to configure the following Massif options:
- profile heap specifies whether or not heap profiling should be performed. This option is enabled by default.
- profile stack specifies whether or not stack profiling should be performed. This option is disabled by default, as it significantly slows down Massif.
- heap allocation threshold specifies the significance threshold for heap allocations (as a percentage). Allocation tree entries that account for less than this will be aggregated. The default value for this option is 1%.
- time unit specifies what time unit should be used for the profile. The possible values for this are instructions (default), milliseconds, or bytes.
- max snapshots specifies the maximum number of snapshots that the profile should take. The default value for this is 100.
- The allocated functions field is identical to the Massif command-line option
--alloc-fn=, which allows you to specify any function to be treated as a heap allocation function. To add or remove functions to this field, use the New or Remove buttons. You can specify multiple functions in this field.
- administrative bytes per block is identical to the Massif command-line option
--heap-admin=, which specifies the number of administrative bytes (per block) to use if heap profiling is enabled. The default value for this option is 8.
- allocation tree depth is identical to the Massif command-line option
--depth=, which specifies the maximum depth of the allocation trees recorded for detailed snapshots. The default value for this option is 30.
- allocation peak inaccuracy is identical to the Massif command-line option
--peak-inaccuracy=. Massif records a peak only when the global memory allocation size exceeds the previous peak by the specified percentage value of this option. The default value for this option is 1.
- detailed snapshot frequency specifies the frequency at which Massif should take detailed snapshots. The default value for this option is 10; to specify that each snapshot should be detailed, set this option to 1.
- minimum heap block alignment specifies the minimum alignment (i.e. size) of heap blocks.
For more information about each Massif option in the Valgrind plug-in for Eclipse, refer to
Cachegrind performs cache and branching profiling. A Cachegrind profile run measures the number of cache misses and branch mispredictions performed by an application. Cachegrind collects the following statistics:
- L1 instruction cache reads and misses
- L1 data cache reads and read misses, writes, and write misses
- L2 unified cache reads and read misses, writes and write misses
- Conditional branches and mispredicted conditional branches
- Indirect branches and mispredicted indirect brances
Analyzing Cachegrind Profile Results
The results of a Cachegrind profile run are displayed in the Valgrind view. These results show Cachegrind's cache/branch data in different levels of granularity. Double-clicking on any file, function, or line will open the corresponding source file and place the cursor on the appropriate location (if the source can be resolved).
Configuring a Cachegrind Profile Run
The Cachegrind Options tab allows you to configure the following Cachegrind options:
- Profile Cache Accesses/Misses is identical to the Cachegrind command-line option
--cache-sim=, which specifies whether or not to collect cache accesses and miss counts. This option is enabled by default.
- Profile Branch Instructions/Mispredictions is identical to the Cachegrind command-line option
--branch-sim=, wich specifies whether or not to collect branch instruction and misprediction counts. This option is disabled by default.
- The options inside the Manually Set Cache Specifications box allow you to configure the size, associativity, and line size of the following caches:
- l1 (level 1 instruction) cache
- D1 (level 1 data) cache
- L2 (level 2) cache
For more information about each Cachegrind option in the Valgrind plug-in for Eclipse, refer to
If the profiled application contains any errors or if any Valgrind options you selected result in an error, these errors will be reported in the Core Messages pane of the Valgrind View. The Memcheck tool outputs all of its profile results on this pane.
Profiling Child Processes
To profile child processes during a profile run, enable the Trace children on exec option in the General Options tab (of the Valgrind Options tab). Memcheck's results will list the process ID (PID) for each reported error in the Valgrind view.
Cachegrind will report each PID separately as a top-level element in its output tree.