7.4.3. KOKKOS package
Kokkos is a templated C++ library that provides abstractions to allow a single implementation of an application kernel (e.g. a pair style) to run efficiently on different kinds of hardware, such as GPUs, Intel Xeon Phis, or many-core CPUs. Kokkos maps the C++ kernel onto different backend languages such as CUDA, OpenMP, or Pthreads. The Kokkos library also provides data abstractions to adjust (at compile time) the memory layout of data structures like 2d and 3d arrays to optimize performance on different hardware. For more information on Kokkos, see Github. Kokkos is part of Trilinos. The Kokkos library was written primarily by Carter Edwards, Christian Trott, and Dan Sunderland (all Sandia).
The LAMMPS KOKKOS package contains versions of pair, fix, and atom styles that use data structures and macros provided by the Kokkos library, which is included with LAMMPS in /lib/kokkos. The KOKKOS package was developed primarily by Christian Trott (Sandia) and Stan Moore (Sandia) with contributions of various styles by others, including Sikandar Mashayak (UIUC), Ray Shan (Sandia), and Dan Ibanez (Sandia). For more information on developing using Kokkos abstractions see the Kokkos programmers’ guide at /lib/kokkos/doc/Kokkos_PG.pdf.
Kokkos currently provides support for 3 modes of execution (per MPI task). These are Serial (MPI-only for CPUs and Intel Phi), OpenMP (threading for many-core CPUs and Intel Phi), and CUDA (for NVIDIA GPUs). You choose the mode at build time to produce an executable compatible with specific hardware.
Kokkos support within LAMMPS must be built with a C++11 compatible compiler. This means GCC version 4.7.2 or later, Intel 14.0.4 or later, or Clang 3.5.2 or later is required.
To build with Kokkos support for NVIDIA GPUs, NVIDIA CUDA software version 7.5 or later must be installed on your system. See the discussion for the GPU package for details of how to check and do this.
Kokkos with CUDA currently implicitly assumes, that the MPI library is CUDA-aware and has support for GPU-direct. This is not always the case, especially when using pre-compiled MPI libraries provided by a Linux distribution. This is not a problem when using only a single GPU and a single MPI rank on a desktop. When running with multiple MPI ranks, you may see segmentation faults without GPU-direct support. These can be avoided by adding the flags -pk kokkos gpu/direct off to the LAMMPS command line or by using the command package kokkos gpu/direct off in the input file.
Building LAMMPS with the KOKKOS package:
See the Build extras doc page for instructions.
Running LAMMPS with the KOKKOS package:
All Kokkos operations occur within the context of an individual MPI task running on a single node of the machine. The total number of MPI tasks used by LAMMPS (one or multiple per compute node) is set in the usual manner via the mpirun or mpiexec commands, and is independent of Kokkos. E.g. the mpirun command in OpenMPI does this via its -np and -npernode switches. Ditto for MPICH via -np and -ppn.
Running on a multi-core CPU:
Here is a quick overview of how to use the KOKKOS package for CPU acceleration, assuming one or more 16-core nodes.
mpirun -np 16 lmp_kokkos_mpi_only -k on -sf kk -in in.lj # 1 node, 16 MPI tasks/node, no multi-threading mpirun -np 2 -ppn 1 lmp_kokkos_omp -k on t 16 -sf kk -in in.lj # 2 nodes, 1 MPI task/node, 16 threads/task mpirun -np 2 lmp_kokkos_omp -k on t 8 -sf kk -in in.lj # 1 node, 2 MPI tasks/node, 8 threads/task mpirun -np 32 -ppn 4 lmp_kokkos_omp -k on t 4 -sf kk -in in.lj # 8 nodes, 4 MPI tasks/node, 4 threads/task
To run using the KOKKOS package, use the “-k on”, “-sf kk” and “-pk kokkos” command-line switches in your mpirun command. You must use the “-k on” command-line switch to enable the KOKKOS package. It takes additional arguments for hardware settings appropriate to your system. For OpenMP use:
-k on t Nt
The “t Nt” option specifies how many OpenMP threads per MPI task to use with a node. The default is Nt = 1, which is MPI-only mode. Note that the product of MPI tasks * OpenMP threads/task should not exceed the physical number of cores (on a node), otherwise performance will suffer. If hyperthreading is enabled, then the product of MPI tasks * OpenMP threads/task should not exceed the physical number of cores * hardware threads. The “-k on” switch also issues a “package kokkos” command (with no additional arguments) which sets various KOKKOS options to default values, as discussed on the package command doc page.
The “-sf kk” command-line switch will automatically append the “/kk” suffix to styles that support it. In this manner no modification to the input script is needed. Alternatively, one can run with the KOKKOS package by editing the input script as described below.
The default for the package kokkos command is to use “full” neighbor lists and set the Newton flag to “off” for both pairwise and bonded interactions. However, when running on CPUs, it will typically be faster to use “half” neighbor lists and set the Newton flag to “on”, just as is the case for non-accelerated pair styles. It can also be faster to use non-threaded communication. Use the “-pk kokkos” command-line switch to change the default package kokkos options. See its doc page for details and default settings. Experimenting with its options can provide a speed-up for specific calculations. For example:
mpirun -np 16 lmp_kokkos_mpi_only -k on -sf kk -pk kokkos newton on neigh half comm no -in in.lj # Newton on, Half neighbor list, non-threaded comm
If the newton command is used in the input script, it can also override the Newton flag defaults.
Core and Thread Affinity:
When using multi-threading, it is important for performance to bind both MPI tasks to physical cores, and threads to physical cores, so they do not migrate during a simulation.
If you are not certain MPI tasks are being bound (check the defaults for your MPI installation), binding can be forced with these flags:
OpenMPI 1.8: mpirun -np 2 --bind-to socket --map-by socket ./lmp_openmpi ... Mvapich2 2.0: mpiexec -np 2 --bind-to socket --map-by socket ./lmp_mvapich ...
For binding threads with KOKKOS OpenMP, use thread affinity environment variables to force binding. With OpenMP 3.1 (gcc 4.7 or later, intel 12 or later) setting the environment variable OMP_PROC_BIND=true should be sufficient. In general, for best performance with OpenMP 4.0 or better set OMP_PROC_BIND=spread and OMP_PLACES=threads. For binding threads with the KOKKOS pthreads option, compile LAMMPS the KOKKOS HWLOC=yes option as described below.
Running on Knight’s Landing (KNL) Intel Xeon Phi:
Here is a quick overview of how to use the KOKKOS package for the Intel Knight’s Landing (KNL) Xeon Phi:
KNL Intel Phi chips have 68 physical cores. Typically 1 to 4 cores are reserved for the OS, and only 64 or 66 cores are used. Each core has 4 hyperthreads,so there are effectively N = 256 (4*64) or N = 264 (4*66) cores to run on. The product of MPI tasks * OpenMP threads/task should not exceed this limit, otherwise performance will suffer. Note that with the KOKKOS package you do not need to specify how many KNLs there are per node; each KNL is simply treated as running some number of MPI tasks.
Examples of mpirun commands that follow these rules are shown below.
Intel KNL node with 68 cores (272 threads/node via 4x hardware threading): mpirun -np 64 lmp_kokkos_phi -k on t 4 -sf kk -in in.lj # 1 node, 64 MPI tasks/node, 4 threads/task mpirun -np 66 lmp_kokkos_phi -k on t 4 -sf kk -in in.lj # 1 node, 66 MPI tasks/node, 4 threads/task mpirun -np 32 lmp_kokkos_phi -k on t 8 -sf kk -in in.lj # 1 node, 32 MPI tasks/node, 8 threads/task mpirun -np 512 -ppn 64 lmp_kokkos_phi -k on t 4 -sf kk -in in.lj # 8 nodes, 64 MPI tasks/node, 4 threads/task
The -np setting of the mpirun command sets the number of MPI tasks/node. The “-k on t Nt” command-line switch sets the number of threads/task as Nt. The product of these two values should be N, i.e. 256 or 264.
The default for the package kokkos command is to use “full” neighbor lists and set the Newton flag to “off” for both pairwise and bonded interactions. When running on KNL, this will typically be best for pair-wise potentials. For manybody potentials, using “half” neighbor lists and setting the Newton flag to “on” may be faster. It can also be faster to use non-threaded communication. Use the “-pk kokkos” command-line switch to change the default package kokkos options. See its doc page for details and default settings. Experimenting with its options can provide a speed-up for specific calculations. For example:
mpirun -np 64 lmp_kokkos_phi -k on t 4 -sf kk -pk kokkos comm no -in in.lj # Newton off, full neighbor list, non-threaded comm mpirun -np 64 lmp_kokkos_phi -k on t 4 -sf kk -pk kokkos newton on neigh half comm no -in in.reax # Newton on, half neighbor list, non-threaded comm
MPI tasks and threads should be bound to cores as described above for CPUs.
To build with Kokkos support for Intel Xeon Phi coprocessors such as Knight’s Corner (KNC), your system must be configured to use them in “native” mode, not “offload” mode like the USER-INTEL package supports.
Running on GPUs:
Use the “-k” command-line switch to specify the number of GPUs per node. Typically the -np setting of the mpirun command should set the number of MPI tasks/node to be equal to the number of physical GPUs on the node. You can assign multiple MPI tasks to the same GPU with the KOKKOS package, but this is usually only faster if significant portions of the input script have not been ported to use Kokkos. Using CUDA MPS is recommended in this scenario. Using a CUDA-aware MPI library with support for GPU-direct is highly recommended. GPU-direct use can be avoided by using -pk kokkos gpu/direct no. As above for multi-core CPUs (and no GPU), if N is the number of physical cores/node, then the number of MPI tasks/node should not exceed N.
-k on g Ng
Here are examples of how to use the KOKKOS package for GPUs, assuming one or more nodes, each with two GPUs:
mpirun -np 2 lmp_kokkos_cuda_openmpi -k on g 2 -sf kk -in in.lj # 1 node, 2 MPI tasks/node, 2 GPUs/node mpirun -np 32 -ppn 2 lmp_kokkos_cuda_openmpi -k on g 2 -sf kk -in in.lj # 16 nodes, 2 MPI tasks/node, 2 GPUs/node (32 GPUs total)
The default for the package kokkos command is to use “full” neighbor lists and set the Newton flag to “off” for both pairwise and bonded interactions, along with threaded communication. When running on Maxwell or Kepler GPUs, this will typically be best. For Pascal GPUs, using “half” neighbor lists and setting the Newton flag to “on” may be faster. For many pair styles, setting the neighbor binsize equal to the ghost atom cutoff will give speedup. Use the “-pk kokkos” command-line switch to change the default package kokkos options. See its doc page for details and default settings. Experimenting with its options can provide a speed-up for specific calculations. For example:
mpirun -np 2 lmp_kokkos_cuda_openmpi -k on g 2 -sf kk -pk kokkos binsize 2.8 -in in.lj # Set binsize = neighbor ghost cutoff mpirun -np 2 lmp_kokkos_cuda_openmpi -k on g 2 -sf kk -pk kokkos newton on neigh half binsize 2.8 -in in.lj # Newton on, half neighborlist, set binsize = neighbor ghost cutoff
For good performance of the KOKKOS package on GPUs, you must have Kepler generation GPUs (or later). The Kokkos library exploits texture cache options not supported by Telsa generation GPUs (or older).
When using a GPU, you will achieve the best performance if your input script does not use fix or compute styles which are not yet Kokkos-enabled. This allows data to stay on the GPU for multiple timesteps, without being copied back to the host CPU. Invoking a non-Kokkos fix or compute, or performing I/O for thermo or dump output will cause data to be copied back to the CPU incurring a performance penalty.
To get an accurate timing breakdown between time spend in pair, kspace, etc., you must set the environment variable CUDA_LAUNCH_BLOCKING=1. However, this will reduce performance and is not recommended for production runs.
Run with the KOKKOS package by editing an input script:
The discussion above for building LAMMPS with the KOKKOS package, the mpirun/mpiexec command, and setting appropriate thread are the same.
You must still use the “-k on” command-line switch to enable the KOKKOS package, and specify its additional arguments for hardware options appropriate to your system, as documented above.
You can use the suffix kk command, or you can explicitly add a “kk” suffix to individual styles in your input script, e.g.
pair_style lj/cut/kk 2.5
Using OpenMP threading and CUDA together (experimental):
With the KOKKOS package, both OpenMP multi-threading and GPUs can be used together in a few special cases. In the Makefile, the KOKKOS_DEVICES variable must include both “Cuda” and “OpenMP”, as is the case for /src/MAKE/OPTIONS/Makefile.kokkos_cuda_mpi
The suffix “/kk” is equivalent to “/kk/device”, and for Kokkos CUDA, using the “-sf kk” in the command line gives the default CUDA version everywhere. However, if the “/kk/host” suffix is added to a specific style in the input script, the Kokkos OpenMP (CPU) version of that specific style will be used instead. Set the number of OpenMP threads as “t Nt” and the number of GPUs as “g Ng”
-k on t Nt g Ng
For example, the command to run with 1 GPU and 8 OpenMP threads is then:
mpiexec -np 1 lmp_kokkos_cuda_openmpi -in in.lj -k on g 1 t 8 -sf kk
Conversely, if the “-sf kk/host” is used in the command line and then the “/kk” or “/kk/device” suffix is added to a specific style in your input script, then only that specific style will run on the GPU while everything else will run on the CPU in OpenMP mode. Note that the execution of the CPU and GPU styles will NOT overlap, except for a special case:
A kspace style and/or molecular topology (bonds, angles, etc.) running on the host CPU can overlap with a pair style running on the GPU. First compile with “–default-stream per-thread” added to CCFLAGS in the Kokkos CUDA Makefile. Then explicitly use the “/kk/host” suffix for kspace and bonds, angles, etc. in the input file and the “kk” suffix (equal to “kk/device”) on the command line. Also make sure the environment variable CUDA_LAUNCH_BLOCKING is not set to “1” so CPU/GPU overlap can occur.
Speed-ups to expect:
The performance of KOKKOS running in different modes is a function of your hardware, which KOKKOS-enable styles are used, and the problem size.
Generally speaking, the following rules of thumb apply:
- When running on CPUs only, with a single thread per MPI task, performance of a KOKKOS style is somewhere between the standard (un-accelerated) styles (MPI-only mode), and those provided by the USER-OMP package. However the difference between all 3 is small (less than 20%).
- When running on CPUs only, with multiple threads per MPI task, performance of a KOKKOS style is a bit slower than the USER-OMP package.
- When running large number of atoms per GPU, KOKKOS is typically faster than the GPU package.
- When running on Intel hardware, KOKKOS is not as fast as the USER-INTEL package, which is optimized for that hardware.
See the Benchmark page of the LAMMPS web site for performance of the KOKKOS package on different hardware.
Advanced Kokkos options:
There are other allowed options when building with the KOKKOS package. As explained on the Build extras doc page, they can be set either as variables on the make command line or in Makefile.machine, or they can be specified as CMake variables. Each takes a value shown below. The default value is listed, which is set in the lib/kokkos/Makefile.kokkos file.
- KOKKOS_DEBUG, values = yes, no, default = no
- KOKKOS_USE_TPLS, values = hwloc, librt, experimental_memkind, default = none
- KOKKOS_CXX_STANDARD, values = c++11, c++1z, default = c++11
- KOKKOS_OPTIONS, values = aggressive_vectorization, disable_profiling, default = none
- KOKKOS_CUDA_OPTIONS, values = force_uvm, use_ldg, rdc, enable_lambda, default = enable_lambda
KOKKOS_USE_TPLS=hwloc binds threads to hardware cores, so they do not migrate during a simulation. KOKKOS_USE_TPLS=hwloc should always be used if running with KOKKOS_DEVICES=Pthreads for pthreads. It is not necessary for KOKKOS_DEVICES=OpenMP for OpenMP, because OpenMP provides alternative methods via environment variables for binding threads to hardware cores. More info on binding threads to cores is given on the Speed omp doc page.
KOKKOS_USE_TPLS=librt enables use of a more accurate timer mechanism on most Unix platforms. This library is not available on all platforms.
KOKKOS_DEBUG is only useful when developing a Kokkos-enabled style within LAMMPS. KOKKOS_DEBUG=yes enables printing of run-time debugging information that can be useful. It also enables runtime bounds checking on Kokkos data structures.
KOKKOS_CXX_STANDARD and KOKKOS_OPTIONS are typically not changed when building LAMMPS.
KOKKOS_CUDA_OPTIONS are additional options for CUDA. The LAMMPS KOKKOS package must be compiled with the enable_lambda option when using GPUs.
Currently, there are no precision options with the KOKKOS package. All compilation and computation is performed in double precision.