RAJA Performance Suite
The RAJA performance suite is designed to explore performance of loop-based computational kernels of the sort found in HPC applications. In particular, it is used to assess, monitor, and compare runtime performance of kernels implemented using RAJA and variants implemented using standard or vendor-supported parallel programming models directly. Each kernel in the suite appears in multiple RAJA and non-RAJA variants using parallel programming models such as OpenMP and CUDA.
The kernels are borrowed from a variety of sources, such as HPC benchmark suites and applications. Kernels are partitioned into "groups" -- each group indicates the origin of its kernels, the algorithm patterns they represent, etc. For example, the "Apps" group contains a collection of kernels extracted from real scientific computing applications, kernels in the "Basic" group are small and simple, but exhibit challenges for compiler optimizations, and so forth.
Table of Contents
- Building the suite
- Running the suite
- Generated output
- Adding kernels and variants
- Contributions
- Authors
- Copyright and Release
Building the suite
To build the suite, you must first obtain a copy of the code by cloning the source repository. For example,
> mkdir RAJA-PERFSUITE
> cd RAJA-PERFSUITE
> git clone --recursive https://github.com/llnl/RAJAPerf.git
> ls
RAJAPerf
The Performance Suite has RAJA and the CMake-based BLT build system as Git submodules. The '--recursive' argument will clone theses submodules int the Performance Suite source code. Note that if you switch to a different branch, you will have to update the submodules; e.g.,
> cd RAJAPerf
> git checkout <some branch name>
> git submodule init
> git submodule update
RAJA and the Performance Suite are built together using the same CMake configuration. For convenience, we include scripts in the 'scripts' directory that invoke associated (CMake cache) configuration files in the 'host-configs' directory that illustrate how to build the code on various platforms at LLNL. Each build script creates a descriptively-named build space directory in the top-level Performance Suite directory and runs CMake with a configuration appropriate for the platform and compilers used. After CMake completes, enter the build directory and type 'make' (or 'make -j' for a parallel build) to build the code. The provided configurations will build RAJA unit tests by default. After the code builds, you can type 'make test' to verify that the RAJA build is working properly. For example,
> ./scripts/blueos_nvcc8.0_clang-coral.sh
> cd build_blueos_nvcc8.0_clang-coral
> make -j
> make test
You can also create your own build directory and run CMake with your own arguments from there; e.g., :
> mkdir my-build
> cd my-build
> cmake <cmake args> ../
> make -j
Running the suite
The suite is run by invoking the executable in the 'bin' directory in the build space. For example, giving it no options:
> ./bin/raja-perf.exe
will run the entire suite (all kernels and variants) in their default configurations.
The suite can be run in a variety of ways by passing options to the executable. For example, you can run subsets of kernels by specifying variants, group, or listing them explicitly. Other configuration options to set problem sizes, number of kernel repetitions, etc. can also be provided. The goal is to build the code once and use scripts or other mechanisms to run the suite in different ways.
Note: most options appear in a long or short form for ease of use.
To see available options along with a brief description of each, pass the '--help' or '-h' option:
> ./bin/raja-perf.exe --help
or
> ./bin/raja-perf.exe -h
Lastly, the program will emit a summary of provided input if it is given something that it does not understand. Hopefully, this will make it easy for users to understand and correct erroneous usage.
Important notes
-
The kernels that use RAJA 'nested' loop constructs will be replaced at some point with new RAJA nested capabilities that are being developed. The new nested constructs are simpler, more flexible, and perform better.
-
The OpenMP target variants of the kernels in the Suite are a work-in-progress. They are incomplete (a few RAJA features must be filled in to make them comparable to other variants).
-
The build system for the Suite needs to be reworked to have the OpenMP target kernel variants run from the same executable as the CUDA variants. Currently, a separate executable
./bin/raja-perf-nolibs.exe
is generated for running OpenMP target variants when they are enabled.
Generated output
Running the suite will generate several output files whose name starts with
the specified file prefix in the specified out put directory. If no such
preferences are provided, files will be located in the current directory
and be named RAJAPerf*
.
Currently, there are up to four files generated:
- Timing -- execution time (sec.) of each loop kernel and variant
- Checksum -- checksum value from results of each loop kernel and variant
- Speedup -- runtime speedup of each loop kernel and variant with respect to reference variant. Reference variant can be set with command line option.
- Figure of Merit (FOM) -- basic statistics about speedup of RAJA variant vs. baseline for each programming model run. PASS/FAIL tolerance can be set with command line option.
The name of each file is indicative of its contents. All files are text files. Other than the checksum file, all are in 'csv' format for easy processing by various tools.
Adding kernels and variants
This section describes how to add new kernels and/or kernel variants to the RAJA Performance Suite. Group modifications are not required unless a new group is added. The information in this section also provides insight into how the performance suite operates.
It is essential that the appropriate targets are updated in the appropriate
CMakeLists.txt
files when files are added to the suite.
Adding a kernel
Adding a new kernel to the suite involves three main steps:
- Add unique kernel ID and unique name to the suite.
- If the kernel is part of a new kernel group, also add a unique group ID and name for the group.
- Implement a kernel class that contains all operations needed to run it, with source files organized as described below.
These steps are described in the following sections.
Add the kernel ID and name
Two key pieces of information identify a kernel: the group in which it
resides and the name of the kernel itself. For concreteness, we describe
how to add a kernel "Foo" that lives in the kernel group "Bar". The files
RAJAPerfSuite.hpp
and RAJAPerfSuite.cpp
define enumeration
values and arrays of string names for the kernels, respectively.
First, add an enumeration value identifier for the kernel, that is unique
among all kernels, in the enum 'KerneID' in the header file RAJAPerfSuite.hpp
:
enum KernelID {
..
Bar_Foo,
..
};
Note: the enumeration value for the kernel is the group name followed by the kernel name, separated by an underscore. It is important to follow this convention so that the kernel works with existing performance suite machinery.
Second, add the kernel name to the array of strings 'KernelNames' in the file
RAJAPerfSuite.cpp
:
static const std::string KernelNames [] =
{
..
std::string("Bar_Foo"),
..
};
Note: the kernel string name is just a string version of the kernel ID. This convention must be followed so that the kernel works with existing performance suite machinery. Also, the values in the KernelID enum and the strings in the KernelNames array must be kept consistent (i.e., same order and matching one-to-one).
Add new group if needed
If a kernel is added as part of a new group of kernels in the suite, a
new value must be added to the 'GroupID' enum in the header file
RAJAPerfSuite.hpp
and an associated group string name must be added to
the 'GroupNames' array of strings in the file RAJAPerfSuite.cpp
. Again,
the enumeration values and items in the string array must be kept
consistent.
Add the kernel class
Each kernel in the suite is implemented in a class whose header and implementation files live in the directory named for the group in which the kernel lives. The kernel class is responsible for implementing all operations needed to execute and record execution timing and result checksum information for each variant of the kernel.
Continuing with our example, we add 'Foo' class header and implementation
files 'Foo.hpp', 'Foo.cpp' (CPU variants), Foo-Cuda.cpp
(CUDA variants),
and Foo-OMPTarget.cpp
(OpenMP target variants) to the 'src/bar' directory.
The class must inherit from the 'KernelBase' base class that defines the
interface for kernels in the suite.
Kernel class header
Here is what the header file for the Foo kernel object may look like:
#ifndef RAJAPerf_Bar_Foo_HXX
#define RAJAPerf_Bar_Foo_HXX
///
/// Foo kernel reference implementation:
///
/// Describe it here...
///
#include "common/KernelBase.hpp"
namespace rajaperf
{
class RunParams; // Forward declaration for ctor arg.
namespace bar
{
class Foo : public KernelBase
{
public:
Foo(const RunParams& params);
~Foo();
void setUp(VariantID vid);
void runKernel(VariantID vid);
void updateChecksum(VariantID vid);
void tearDown(VariantID vid);
void runCudaVariant(VariantID vid);
void runOpenMPTargetVariant(VariantID vid);
private:
// Kernel-specific data (pointers, scalars, etc.) used in kernel...
};
} // end namespace bar
} // end namespace rajaperf
#endif // closing endif for header file include guard
The kernel object header has a uniquely-named header file include guard and the class is nested within the 'rajaperf' and 'bar' namespaces. The constructor takes a reference to a 'RunParams' object, which contains the input parameters for running the suite -- we'll say more about this later. The four methods that take a variant ID argument must be provided as they are pure virtual in the KernelBase class. Their names are descriptive of what they do and we'll provide more details when we describe the class implementation next.
Kernel class implementation
All kernels in the suite follow a similar implementation pattern for consistency and ease of understanding. Here we describe several steps and conventions that must be followed to ensure that all kernels interact with the performance suite machinery in the same way:
- Initialize the 'KernelBase' class object with KernelID, default size, and default repetition count in the
class constructor
. - Implement data allocation and initialization operation for each kernel variant in the
setUp
method. - Implement kernel execution for each variant in the
RunKernel
method. - Compute the checksum for each variant in the
updateChecksum
method. - Deallocate and reset any data that will be allocated and/or initialized in subsequent kernel executions in the
tearDown
method.
Constructor and destructor
It is important to note that there will only be one instance of each kernel class created by the program. Thus, each kernel class constructor and destructor must only perform operations that are non-specific to any kernel variant.
The constructor must pass the kernel ID and RunParams object to the base class 'KernelBase' constructor. The body of the constructor must also call base class methods to set the default size for the iteration space of the kernel (e.g., typically the number of loop iterations, but can be kernel-dependent) and the number of times to repeat (i.e., execute) the kernel with each pass through the suite to generate adequate timing information. These values will be modified based on input parameters to define the actual size and number of reps applied when the suite is run. Here is how this typically looks:
Foo::Foo(const RunParams& params)
: Foo(rajaperf::Bar_Foo, params),
// default initialization of class members
{
setDefaultSize(100000);
setDefaultReps(1000);
}
The class destructor doesn't have any requirements beyond freeing memory owned by the class object as needed.
setUp() method
The 'setUp()' method is responsible for allocating and initializing data necessary to run the kernel for the variant specified by its variant ID argument. For example, a baseline variant may have aligned data allocation to help enable SIMD optimizations, an OpenMP variant may initialize arrays following a pattern of "first touch" based on how memory and threads are mapped to CPU cores, a CUDA variant may initialize data in host memory and copy it into device memory, etc.
It is important to use the same data allocation and initialization operations for RAJA and non-RAJA variants that are related. Also, the state of all input data for the kernel should be the same for all variants so that checksums can be compared at the end of a run.
Note: to simplify these operations and help ensure consistency, there exist
utility methods to allocate, initialize, deallocate, and copy data, and compute
checksums defined in the DataUtils.hpp
CudaDataUtils.hpp
, and
OpenMPTargetDataUtils.hpp
header files in the 'common' directory.
runKernel() method
The 'runKernel()' method executes the kernel for the variant defined by the variant ID argument. The method is also responsible for calling base class methods to start and stop execution timers for the loop variant. A typical kernel execution code section may look like:
void Foo::runKernel(VariantID vid)
{
const Index_type run_reps = getRunReps();
// ...
switch ( vid ) {
case Base_Seq : {
// Declare data for baseline sequential variant of kernel...
startTimer();
for (SampIndex_type irep = 0; irep < run_reps; ++irep) {
// Implementation of kernel variant...
}
stopTimer();
// ...
break;
}
// case statements for other CPU kernel variants....
#if defined(RAJA_ENABLE_TARGET_OPENMP)
case Base_OpenMPTarget :
case RAJA_OpenMPTarget :
{
runOpenMPTargetVariant(vid);
break;
}
#endif
#if defined(RAJA_ENABLE_CUDA)
case Base_CUDA :
case RAJA_CUDA :
{
runCudaVariant(vid);
break;
}
#endif
default : {
std::cout << "\n <kernel-name> : Unknown variant id = " << vid << std::endl;
}
}
}
All kernel implementation files are organized in this way. So following this pattern will keep all new additions consistent.
Note: There are three source files for each kernel: 'Foo.cpp' contains CPU
variants, Foo-Cuda.cpp
contains CUDA variants, and Foo-OMPTarget.cpp
constains OpenMP target variants. The reason for this is that it makes it
easier to apply unique compiler flags to different variants and to manage
compilation and linking issues that arise when some kernel variants are
combined in the same translation unit.
Note: for convenience, we make heavy use of macros to define data declarations and kernel bodies in the suite. This significantly reduces the amount of redundant code required to implement multiple variants of each kernel. The kernel class implementation files in the suite provide many examples of the basic pattern we use.
updateChecksum() method
The 'updateChecksum()' method is responsible for adding the checksum for the current kernel (based on the data the kernel computes) to the checksum value for the variant of the kernel just executed, which is held in the KernelBase base class object.
It is important that the checksum be computed in the same way for each variant of the kernel so that checksums for different variants can be compared to help identify differences, and potentially errors, in implementations, compiler optimizations, programming model execution, etc.
Note: to simplify checksum computations and help ensure consistency, there
are methods to compute checksums defined in the DataUtils.hpp
header file
in the 'common' directory.
tearDown() method
The 'tearDown()' method free and/or reset all kernel data that is allocated and/or initialized in the 'setUp' method execution to prepare for other kernel variants run subsequently.
Add object construction operation
The 'Executor' class object is responsible for creating kernel objects
for the kernels to be run based on the suite input options. To ensure a new
kernel object will be created properly, add a call to its class constructor
based on its 'KernelID' in the 'getKernelObject()' method in the
RAJAPerfSuite.cpp
file.
Adding a variant
Each variant in the RAJA Performance Suite is identified by an enumeration value and a string name. Adding a new variant requires adding these two items similar to adding a kernel as described above.
Add the variant ID and name
First, add an enumeration value identifier for the variant, that is unique
among all variants, in the enum 'VariantID' in the header file
RAJAPerfSuite.hpp
:
enum VariantID {
..
NewVariant,
..
};
Second, add the variant name to the array of strings 'VariantNames' in the file
RAJAPerfSuite.cpp
:
static const std::string VariantNames [] =
{
..
std::string("NewVariant"),
..
};
Note that the variant string name is just a string version of the variant ID. This convention must be followed so that the variant works with existing performance suite machinery. Also, the values in the VariantID enum and the strings in the VariantNames array must be kept consistent (i.e., same order and matching one-to-one).
Add kernel variant implementations
In the classes containing kernels to which the new variant applies, add implementations for the variant in the setup, kernel execution, checksum computation, and teardown methods. These operations are described in earlier sections for adding a new kernel above.
Contributions
The RAJA Performance Suite is intended to remain a work-in-progress, with new kernels and variants added over time. We encourage interested parties to contribute to it so that C++ compiler optimizations and support for programming models like RAJA continue to improve.
The Suite developers follow the GitFlow development model. Folks wishing to contribute to the Suite, should include their work in a feature branch created from the RAJA develop
branch. Then, create a pull request with the develop
branch as the
destination when it is ready to be reviewed. The develop
branch contains the
latest work in RAJA Performance Suite. Periodically, we will merge the
develop branch into the master
branch and tag a new release.
Authors
The RAJA Performance Suite was originally developed by:
- Rich Hornung (hornung1@llnl.gov)
Please see the {RAJA Performance Suite Contributors Page](https://github.com/LLNL/RAJAPerf/graphs/contributors), to see the full list of contributors to the project.
Copyright and Release
Copyright (c) 2017, Lawrence Livermore National Security, LLC.
Produced at the Lawrence Livermore National Laboratory.
All rights reserved.
LLNL-CODE-738930
OCEC-17-159
Unlimited Open Source - BSD Distribution
For release details and restrictions, please read the RELEASE, LICENSE, and NOTICE files, also linked here: