/prk_kernels

This is a set of simple programs that can be used to explore the features of a parallel platform.

Primary LanguageCOtherNOASSERTION

PRK logo.

license GitHub contributors GitHub language count GitHub top language

Overview

This suite contains a number of kernel operations, called Parallel Research Kernels, plus a simple build system intended for a Linux-compatible environment. Most of the code relies on open standard programming models and thus can be executed on many computing systems.

These programs should not be used as benchmarks. They are operations to explore features of a hardware platform, but they do not define fixed problems that can be used to rank systems. Furthermore they have not been optimized for the features of any particular system.

Build Instructions

To build the codes the user needs to make certain changes by editing text files. Assuming the source tree is untarred in directory $PRK, the following file needs to be copied to $PRK/common/make.defs and edited.

$PRK/common/make.defs.in -- This file specifies the names of the C compiler (CC), and of the MPI (Message Passing Interface) compiler MPICC or compile script. If MPI is not going to be used, the user can ignore the value of MPICC. The compilers should already be in your path. That is, if you define CC=icc, then typing which icc should show a valid path where that compiler is installed. Special instructions for building and running codes using Charm++, Grappa, OpenSHMEM, or Fine-Grain MPI are in README.special.

We provide examples of working examples for a number of programming environments. Some of these are tested more than others. If you are looking for the simplest option, try make.defs.gcc.

File (in ./common/) Environment
make.defs.cray Cray toolchain (rarely tested).
make.defs.cuda GCC with the CUDA compiler (only used in C++/CUDA implementation).
make.defs.gcc GCC compiler toolchain, which supports essentially all implementations (tested often).
make.defs.freebsd FreeBSD (rarely tested).
make.defs.ibmbg IBM Blue Gene/Q compiler toolchain (deprecated).
make.defs.ibmp9nv IBM compilers for POWER9 and NVIDIA Volta platforms (rarely tested).
make.defs.intel Intel Parallel Studio toolchain, which supports most implementations (tested often).
make.defs.llvm LLVM compiler toolchain, which supports most implementations (tested often).
make.defs.musl GCC compiler toolchain with MUSL as the C standard library, which was required to use C11 threads.
make.defs.nvhpc NVIDIA HPC SDK, which supports most implementations (tested often).
make.defs.oneapi Intel oneAPI.
make.defs.pgi PGI compiler toolchain (infrequently tested).
make.defs.hip HIP compiler toolchain (infrequently tested).

Some of the C++ implementations require you to install Boost, RAJA, Kokkos, Parallel STL, respectively, and then modify make.defs appropriately. Please see the documentation in the documentation (doc) subdirectory.

You can refer to the travis subdirectory for install scripts that can be readily modified to install any of the dependencies in your local environment.

Supported Programming Models

The suite of kernels currently has complete parallel implementations in OpenMP, MPI, Adaptive MPI and Fine-Grain MPI. There is also a SERIAL reference implementation.

The suite is currently being extended to include Charm++, MPI+OpenMP, OpenSHMEM, UPC, and Grappa, Fortran with coarrays, as well as three new variations of MPI:

  1. MPI with one-sided communications (MPIRMA)
  2. MPI with direct use of shared memory inside coherency domains (MPISHM)
  3. MPI with OpenMP inside coherency domains (MPIOPENMP) These extensions are not yet complete.

More recently, we have implemented many single-node programming models in modern languages.

Modern C++

y = yes

i = in-progress, incomplete, incorrect, or incredibly slow

f = see footnotes

Parallelism p2p stencil transpose nstream sparse dgemm PIC
None y y y y y y y
C++11 threads, async y
OpenMP y y y y
OpenMP tasks y y y y
OpenMP target y y y y
OpenCL 1.x i y y y
SYCL i y y y y y
Boost.Compute y
Parallel STL y y y y
Thrust i y
TBB y y y y
Kokkos y y y y
RAJA y y y y
CUDA i y y y
CUBLAS y y y
HIP i y y y
HIPBLAS y y y
CBLAS y y
OpenACC y
MPI (RMA) y

Modern C

Parallelism p2p stencil transpose nstream sparse
None y y y y
C11 threads y
OpenMP y y y y
OpenMP tasks y y y y
OpenMP target y y y y
Cilk y y
ISPC y
MPI y
PETSc i y

There are versions of nstream with OpenMP that support memory allocation using mmap and memkind, which can be used for testing novel memory systems, including persistent memory.

Modern Fortran

Parallelism p2p stencil transpose nstream sparse dgemm
None y y y y y
Intrinsics y y y
coarrays y y y
Global Arrays y y
OpenMP y y y y y
OpenMP tasks y y y y
OpenMP target y y y y
OpenACC y y y

By intrinsics, we mean the language built-in features, such as colon notation or the TRANSPOSE intrinsic. We use DO CONCURRENT in a few places.

Other languages

x = externally supported (in the Chapel repo)

Parallelism p2p stencil transpose nstream sparse dgemm
Python 3 y y y y y y
Python 3 w/ Numpy y y y y y y
Python 3 w/ mpi4py y y y
Julia y y y
Octave (Matlab) y y y
Rust y y y
Go y y y
C# y y
Chapel x x x
Java y y y y
Lua y

Global make

Please run make help in the top directory for the latest information.

To build all available kernels of a certain version, type in the root directory:

Command Effect
make all builds all kernels.
make allserial builds all serial kernels.
make allopenmp builds all OpenMP kernels.
make allmpi builds all conventional two-sided MPI kernels.
make allmpi1 builds all MPI kernels.
make allfgmpi builds all Fine-Grain MPI kernels.
make allampi builds all Adaptive MPI kernels.
make allmpiopenmp builds all hybrid MPI+OpenMP kernels.
make allmpirma builds all MPI-3 kernels with one-sided communications.
make allmpishm builds all kernels with MPI-3 shared memory.
make allshmem builds all OpenSHMEM kernels.
make allupc builds all Unified Parallel C (UPC) kernels.
make allcharm++ builds all Charm++ kernels.
make allgrappa builds all Grappa kernels.
make allfortran builds all Fortran kernels.
make allc1x builds all C99/C11 kernels.
make allcxx builds all C++11 kernels.

The global make process uses a single set of optimization flags for all kernels. For more control, the user should consider individual makes (see below), carefully choosing the right parameters in each Makefile. If a a single set of optimization flags different from the default is desired, the command line can be adjusted: make all<version> default_opt_flags=<list of optimization flags>

The global make process uses some defaults for the Branch kernel (see Makefile in that directory). These can be overridden by adjusting the command line: make all<version> matrix_rank=<n> number_of_functions=<m> Note that no new values for matrix_rank or number_of_functions will be used unless a make veryclean has been issued.

Individual make

Descend into the desired sub-tree and cd to the kernel(s) of interest. Each kernel has its own Makefile. There are a number of parameters that determine the behavior of the kernel that need to be known at compile time. These are explained succinctly in the Makefile itself. Edit the Makefile to activate certain parameters, and/or to set their values.

Typing make without parameters in each leaf directory will prompt the user for the correct parameter syntax. Once the code has been built, typing the name of the executable without any parameters will prompt the user for the correct parameter syntax.

Running test suite

After the desired kernels have been built, they can be tested by executing scripts in the 'scripts' subdirectory from the root of the kernels package. Currently two types of run scripts are supported. scripts/small: tests only very small examples that should complete in just a few seconds. This merely tests functionality of kernels and installed runtimes scripts/wide: tests examples that will take up most memory on a single node with 64 GB of memory.

Only a few parameters can be changed globally; for rigorous testing, the user should run each kernel individually, carefully choosing the right parameters. This may involve editing the individual Makefiles and rerunning the kernels.

Example build and runs

make all default_opt_flags="-O2" "matrix_rank=7" "number_of_functions=200" 
./scripts/small/runopenmp
./scripts/small/runmpi1
./scripts/wide/runserial
./scripts/small/runcharm++
./scripts/wide/runmpiopenmp

To exercise all kernels, type

./scripts/small/runall
./scripts/wide/runall

Quality Control

We have a rather massive test matrix running in Travis CI. Unfortunately, the Travis CI environment may vary with time and occasionally differs from what we are running locally, which makes debugging tricky. If the status of the project is not passing, please inspect the details, because this may not be an indication of an issue with our project, but rather something in Travis CI.

License

See COPYING for licensing information.

Note on stream

Note that while our nstream operations are based on the well known STREAM benchmark by John D. McCalpin, we modified the source code and do not follow the run-rules associated with this benchmark. Hence, according to the rules defined in the STREAM license (see clause 3b), you must never report the results of our nstream operations as official "STREAM Benchmark" results. The results must be clearly labled whenever they are published. Examples of proper labelling include:

  "tuned STREAM benchmark results" 
  "based on a variant of the STREAM benchmark code" 

Other comparable, clear, and reasonable labelling is acceptable.