Hotspot - the Linux perf GUI for performance analysis
This project is a KDAB R&D effort to create a standalone
GUI for performance data. As the first goal, we want to provide a UI like
KCachegrind around Linux perf. Looking ahead, we intend to support various other
performance data formats under this umbrella.
Here are some screenshots showing the most important features of hotspot in action:
Visualize Data
The main feature of hotspot is visualizing a perf.data file graphically.
Time Line
The time line allows filtering the results by time, process or thread. The data views will update accordingly.
Record Data
You can also launch perf from hotspot, to profile a newly started application
or to attach to already running process(es). Do take the caveats below
into account though.
Building Hotspot
Building hotspot from source gives you the latest and greatest, but you'll have to make sure all its dependencies are available. Most users should probably install hotspot from the distro package manager or as an AppImage.
Required Dependencies
As of now, you will need the following dependencies to build this project:
CMake 3.1.0 or higher
any C++11 enabled compiler
Qt 5.15.0 or higher
libelf
libelfutils
gettext
extra-cmake-modules
KDE Frameworks 5 (packages are usually called libkf5-*-devel):
git clone --recurse-submodules https://github.com/KDAB/hotspot.git
mkdir build-hotspot
cd build-hotspot
cmake ../hotspot
make
# now you can run hotspot from the build folder:
./bin/hotspot
# or `make install` it and launch it from your $PATH
If you want to use KAuth, you need to add -DCMAKE_INSTALL_PREFIX=/usr/ to the cmake call. Otherwise KAuth won't work.
If you need help building this project for your platform, contact us for help.
Getting Hotspot
Note: Hotspot is not yet packaged on all Linux distributions. In such cases, or when you want to use the latest version, please use the AppImage which will work on any recent Linux distro just fine.
Head over to our download page and download the latest AppImage release and just run it.
Please use the latest Continuous release to get the most recent version. If it doesn't work, please report a bug
and test the latest stable version.
Note: Your system libraries or preferences are not altered. In case you'd like to remove Hotspot again, simply delete the downloaded file. Learn more about AppImage here.
Using
First of all, record some data with perf. To get backtraces, you will need to enable the dwarf callgraph mode:
perf record --call-graph dwarf <your application>
...
[ perf record: Woken up 58 times to write data ]
[ perf record: Captured and wrote 14.874 MB perf.data (1865 samples) ]
Now, if you have hotspot available on the same machine, all you need to do is launch it.
It will automatically open the perf.data file in the current directory (similar to perf report).
Alternatively, you can specify the path to the data file on the console:
hotspot /path/to/perf.data
Off-CPU Profiling
Hotspot supports a very powerful way of doing wait-time analysis, or off-CPU profiling.
This analysis is based on kernel tracepoints in the linux scheduler. By recording that
data, we can find the time delta during which a thread was not running on the CPU, but
instead was off-CPU. There can be multiple reasons for that, all of which can be found
using this technique:
synchronous I/O, e.g. via read() or write()
page faults, e.g. when accessing mmap()'ed file data
calls to nanosleep() or yield()
lock contention via futex() etc.
preemption
and probably many more
By leveraging kernel trace points in the scheduler, the overhead is pretty manageable
and we only pay a price, when the process is actually getting switched out. Most notably
we do not pay a price when e.g. a mutex lock operation can be handled directly in
user-space.
To do off-CPU analysis with hotspot, you need to record the data with a very specific
command:
$ perf record \
-e cycles \ # on-CPU profiling
-e sched:sched_switch --switch-events \ # off-CPU profiling
--sample-cpu \ # track on which core code is executed
-m 8M \ # reduce chance of event loss
--aio -z \ # reduce disk-I/O overhead and data size
--call-graph dwarf \ # we definitely want backtraces
<your application>
Alternatively, you can use the off-CPU check box in hotspot's integrated record page.
During the analysis, you can then switch between the "cycles" cost view for on-CPU data
to the "off-CPU time" cost view for wait-time analysis. Often, you will want to change
between both, e.g. to find places in your code which may require further parallelization
(see also Amdahl's law).
The "sched:sched_switch" cost will also be shown to you. But in my opinion that is less
useful, as it only indicates the number of scheduler switches. The length of the time
inbetween is often way more interesting to me - and that's what is shown to you in the
"off-CPU time" metric.
Embedded Systems
If you are recording on an embedded system, you will want to analyze the data on your
development machine with hotspot. To do so, make sure your sysroot contains the debug
information required for unwinding (see below). Then record the data on your embedded
system:
embedded$ perf record --call-graph dwarf <your application>
...
[ perf record: Woken up 58 times to write data ]
[ perf record: Captured and wrote 14.874 MB perf.data (1865 samples) ]
embedded$ cp /proc/kallsyms /tmp/kallsyms # make pseudo-file a real file
It's OK if your embedded machine is using a different platform than your host. On your
host, do the following steps then to analyze the data:
The perf.data file format is not self-contained. To analyze it, you need access
to the executables and libraries of the profiled process, together with debug symbols.
This makes it unwieldy to share such files across machines, e.g. to get the help from
a colleague to investigate a performance issue, or for bug reporting purposes.
Hotspot allows you to export the analyzed data, which is then fully self-contained.
This feature is accessible via the "File > Save As" menu action. The data is then
saved in a self-contained *.perfparser file. To import the data into hotspot again,
just open that file directly in place of the original perf.data file.
Note: The file format is not yet stable. Meaning data exported by one version
of hotspot can only be read back in by the same version. This problem will be
resolved in the future, as time permits.
tracepoints
hotspot currently only shows the name of the tracepoints in the timeline.
Known Issues
If anything breaks in the above and the output is less usable than perf report, please report an issue on GitHub.
That said, there are some known issues that people may trip over:
Broken Backtraces
Unwinding the stack to produce a backtrace is a dark art and can go wrong in many ways.
Hotspot relies on perfparser (see below), which in turn relies on libdw from elfutils
to unwind the stack. This works quite well most of the time, but still can go wrong. Most
notably, unwinding will fail when:
an ELF file (i.e. executable or library) referenced by the perf.data file is missing
to fix this, try to use one of the following CLI arguments to let hotspot know where to look for the ELF files:
--debugPaths <paths>: Use this when you have split debug files in non-standard locations
--extraLibPaths <paths>: Use this when you have moved libraries to some other location since recording
--appPath <paths>: This is kind of a combination of the above two fields. The path is traversed recursively, looking for debug files and libraries.
--sysroot <path>: Use this when you try to inspect a data file recorded on an embedded platform
an ELF file is missing debug information
to fix this, install the debug package from your distro
or compile the code in "release with debug" mode, i.e. ensure your compiler is invoked with something like -O2 -g. You will have to repeat the perf record step
potentially both of the above is not an option for you, e.g. when the library is closed source and supplied by a thirdparty vendor. If that is the case,
you may be lucky and the library contains frame pointers. If so, then try to build elfutils from current git master (you want commit a55df2c1, which should be part of 0.170).
This version of elfutils will try to fallback to the frame pointer for unwinding, when the debug information is missing.
your call stacks are too deep
by default, perf record only copies a part of the stack to the data file. This can lead to issues with very deep call stacks, which will be cut off at some point.
This issue will break the top-down call trees in hotspot, as visualized in the Top-Down view or the Flame Graph. To fix this, you can try to increase the stack dump size, i.e.:
perf record --call-graph dwarf,32768
Note that this can dramatically increase the size of the perf.data files - use it with care. Also have a look at man perf record.
For some scenarios, recursive function calls simply fail to be unwound. See also #93
debuginfod
hotspot supports downloading debug symbols via debuginfod.
This can be enabled by either adding download urls in the settings or launching hotspot with DEBUGINFOD_URLS defined in the environment.
Missing Features
Compared to perf report, hotspot misses a lot of features. Some of these are planned to be resolved
in the future. Others may potentially never get implemented. But be aware that the following features
are not available in hotspot currently:
annotate: the caller/callee view shows cost attributed to individual source lines. A basic annotation view like perf annotate on the source level, is available using the Disassembly feature.
the columns in the tables are currently hardcoded, while potentially a user may want to change this to show e.g. cost per-process or thread and so forth
many of the more advanced features, such as --itrace, --mem-mode, --branch-stack and --branch-history, are unsupported
Recording with perf without super user rights
It is not a good idea to launch hotspot with sudo or as root user. See e.g.
Editing Files As Root
for an article on that matter. Issue #83 is
also relevant in this contact.
But without superuser rights, you may see error messages such as the following
when using hotspot's record feature:
You may not have permission to collect stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid:
-1 - Not paranoid at all
0 - Disallow raw tracepoint access for unpriv
1 - Disallow cpu events for unpriv
2 - Disallow kernel profiling for unpriv
To workaround this limitation, hotspot can temporarily elevate the perf privileges.
This is achieved by applying these steps,
bundled into a script that is run via pkexec, kdesudo or kdesu.
The resulting elevated privileges are also required for kernel tracing in general and Off-CPU profiling in particular.
Export File Format
The current data export is limited to a format that can only be read back by hotspot of the same
version. This makes interop with other visualization tools quasi impossible. This is known and
will get improved in the future. Most notably support for export to web viewers such as
perfetto or the Mozilla profiler is
planned but not yet implemented. Patches welcome!
Qt Creator
This project leverages the excellent perfparser utility created by The Qt Company
for their Qt Creator IDE. If you are already using Qt Creator, consider leveraging
its integrated CPU Usage Analyzer.
License
Hotspot is the GPL v2+. See LICENSE.GPL.txt for more information, or contact
[email protected] if any conditions of this licensing are not clear to you.
请发表评论