Caliper is a program instrumentation and performance measurement framework. It is designed as a performance analysis toolbox in a library, allowing one to bake performance analysis capabilities directly into applications and activate them at runtime. Caliper is primarily aimed at HPC applications, but works for any C/C++/Fortran program on Unix/Linux.
Caliper's data collection mechanisms and source-code annotation API support a variety of performance engineering use cases, such as performance profiling, tracing, monitoring, and auto-tuning.
Features include:
- Low-overhead source-code annotation API
- Flexible key:value data model: capture application-specific features for performance analysis
- Fully threadsafe implementation, support for parallel programming models like MPI
- Synchronous (event-based) and asynchronous (sampling) performance data collection
- Trace and profile recording
- Connection to third-party tools, e.g. NVidia NVProf or Intel(R) VTune(tm)
- Measurement and profiling functionality such as timers, PAPI hardware counters, and Linux perf_events
- Memory allocation annotations: associate performance measurements with named memory regions
Extensive documentation is available here: https://llnl.github.io/Caliper/
Usage examples of the C++ and C annotation interfaces are provided in
the examples/apps directory.
See the "Getting started" section below for a brief tutorial.
Building and installing Caliper requires cmake 3.1+ and a current C++11-compatible Compiler. Clone Caliper from github and proceed as follows:
$ git clone https://github.com/LLNL/Caliper.git
$ cd Caliper
$ mkdir build && cd build
$ cmake -DCMAKE_INSTALL_PREFIX=<path to install location> \
-DCMAKE_C_COMPILER=<path to c-compiler> \
-DCMAKE_CXX_COMPILER=<path to c++-compiler> \
..
$ make
$ make install
See the "Build and install" section in the documentation for further information.
Typically, we integrate Caliper into a program by marking source-code sections of interest with descriptive annotations. Performance measurements, trace or profile collection, and reporting functionality can then be enabled with runtime configuration options. Alternatively, third-party tools can connect to Caliper and access information provided by the source-code annotations.
Caliper's source-code annotation API fulfills two purposes: First, it lets us associate performance measurements with user-defined, high-level context information. Second, we can trigger user-defined actions at the instrumentation points, e.g. to measure the time spent in individual regions. Measurement actions can be defined at runtime and are disabled by default; generally, the source-code annotations are lightweight enough to be left in production code.
The annotation APIs are available for C, C++, and Fortran. There are high-level annotation macros for common scenarios such as marking functions, loops, or sections of source-code. In addition, users can export arbitrary key:value pairs to express application-specific concepts.
The following example marks "initialization" and "main loop" phases in a C++ code, and exports the main loop's current iteration counter using the high-level annotation macros:
#include <caliper/cali.h>
int main(int argc, char* argv[])
{
// Mark this function
CALI_CXX_MARK_FUNCTION;
// Mark the "intialization" phase
CALI_MARK_BEGIN("initialization");
int count = 4;
double t = 0.0, delta_t = 1e-6;
CALI_MARK_END("initialization");
// Mark the loop
CALI_CXX_MARK_LOOP_BEGIN(mainloop, "main loop");
for (int i = 0; i < count; ++i) {
// Mark each loop iteration
CALI_CXX_MARK_LOOP_ITERATION(mainloop, i);
// A Caliper snapshot taken at this point will contain
// { "function"="main", "loop"="main loop", "iteration#main loop"=<i> }
// ...
}
CALI_CXX_MARK_LOOP_END(mainloop);
}To use Caliper, add annotation statements to your program and link it against the Caliper library. Programs must be linked with the Caliper runtime (libcaliper.so), as shown in the example link command:
g++ -o app app.o -L<path to caliper installation>/lib64 -lcaliper
Caliper's performance measurement and data collection functionality must be enabled and configured at runtime through Caliper's configuration API, configuration files, or environment variables. By default, Caliper will keep track of the current Caliper context provided by the annotation API calls (allowing third-party tools to access program context information), but won't run any performance measurement or data recording on its own.
Generally, collecting performance data with Caliper requires selecting
a combination of Caliper services that implement specific
functionality and configuring them for the task at hand. However, for
some common scenarios, Caliper provides a set of pre-defined
configuration profiles. These profiles can be activated with the
CALI_CONFIG_PROFILE environment variable. For example, the
runtime-report configuration profile prints the total time (in
microseconds) spent in each code path based on the nesting of
annotated code regions:
$ CALI_CONFIG_PROFILE=runtime-report ./examples/apps/cali-basic-annotations
Path sum#time.duration
main 20.000000
main loop 8.000000
init 10.000000
The example shows Caliper output for the runtime-report
configuration profile for the source-code annotation example above.
As another example, the serial-trace configuration profile
configures Caliper to record an event trace of each annotation event:
$ CALI_CONFIG_PROFILE=serial-trace ./examples/apps/cali-basic-annotations
== CALIPER: Registered event trigger service
== CALIPER: Registered recorder service
== CALIPER: Registered timestamp service
== CALIPER: Registered trace service
== CALIPER: Initialized
== CALIPER: Flushing Caliper data
== CALIPER: Trace: Flushed 14 snapshots.
== CALIPER: Recorder: Wrote 71 records.
The trace data is stored in a .cali file in a text-based
Caliper-specific file format. Use the cali-query tool to filter,
aggregate, or print the recorded data. Here, we use cali-query to
print the recorded trace data in a human-readable json format:
$ ls *.cali
171120-181836_40337_7LOlCN5RchWV.cali
$ cali-query 171120-181836_40337_7LOlCN5RchWV.cali -q "SELECT * FORMAT json(pretty)"
[
{
"event.begin#function":"main"
},
{
"event.begin#annotation":"init",
"function":"main"
},
{
"event.end#annotation":"init",
"annotation":"init",
"function":"main",
"time.inclusive.duration":14
},
{
"event.begin#loop":"main loop",
"function":"main"
},
{
"loop":"main loop",
"function":"main",
"event.begin#iteration#main loop":0
},
...
As mentioned earlier, Caliper's performance measurement and data collection functionality is provided by independent building blocks called services, each implementing specific functionality (e.g., tracing, I/O, timing, report formatting, sampling, etc.). The services can be enabled at runtime in any combination. This makes Caliper highly flexible, but the runtime configuration can be complex. Refer to the Caliper documentation to learn more.
Caliper was created by David Boehme, [email protected].
A complete list of contributors is available on GitHub.
Major contributors include:
- Alfredo Gimenez (libpfm support, memory allocation tracking)
- David Poliakoff (GOTCHA support)
To reference Caliper in a publication, please cite the following paper:
- David Boehme, Todd Gamblin, David Beckingsale, Peer-Timo Bremer, Alfredo Gimenez, Matthew LeGendre, Olga Pearce, and Martin Schulz. Caliper: Performance Introspection for HPC Software Stacks. In Supercomputing 2016 (SC16), Salt Lake City, Utah, November 13-18, 2016. LLNL-CONF-699263.
Caliper is released under a BSD license. See the LICENSE file for details.
LLNL-CODE-678900