With MPI + OpenMP / OmpSs-2
Table of contents:
Current and near-future High Performance Computing (HPC) systems consist of thousands of parallel computing nodes, connected by high-bandwidth network interconnections, and in most of the cases, each node leveraging one or more GPU devices.
Moreover, some of the most modern GPU devices, such as NVIDIA Tesla V100, support the Unified Memory, which facilitates the task of application developers. With those devices, users do not have to move or copy the data to/from GPUs, and also, pointers at the host are the same at the device.
For these reasons, developers should take benefit from these GPU resources by trying to offload the most compute-intensive parts of the applications to the available GPUs. In this page, we briefly explain the approaches proposed by the OpenMP and the OmpSs-2 programming models to facilitate the offloading of computation tasks to Unified Memory GPUs. Then, we show a hybrid application that has MPI+OpenMP and MPI+OmpSs-2 variants which offload some computation tasks.
On the one hand, OpenMP provides the target directive, which is the one used for
offloading computation parts of OpenMP programs to the GPUs. It provides multiple
clauses to specify the copy directionality of the data, how the computational workload
is distributed, the data dependencies, etc. The user can annotate a part of the
program using the target directive and without having to program it with
a special programming language. For instance, when offloading a part of the program to
NVIDIA GPUs, the user is not required to provide any CUDA kernel. That part
of the program is handled transparently by the compiler.
On the other hand, OmpSs-2 proposes another approach targeting NVIDIA Unified Memory
devices. CUDA kernels can be annotated as regular tasks, and they can declare the
corresponding data dependencies on the data buffers. When all the dependencies of a
CUDA task are satisfied, the CUDA kernel associated with the task is automatically and
asynchronously offloaded to one of the available GPUs. To use that functionality, the
user only has to allocate the buffers that CUDA kernels will access as Unified Memory buffers
(i.e., using the cudaMallocManaged() function). Additionally, users must annotate the CUDA
tasks with the device(cuda) and ndrange(...) clauses.
An N-Body simulation numerically approximates the evolution of a system of bodies in which each body continuously interacts with every other body. A familiar example is an astrophysical simulation in which each body represents a galaxy or an individual star, and the bodies attract each other through the gravitational force.
N-Body simulation arises in many other computational science problems as well. For example, protein folding is studied using N-body simulation to calculate electrostatic and Van der Waals forces. Turbulent fluid flow simulation and global illumination computation in computer graphics are other examples of problems that use N-Body simulation.
Users can clone or download this example from the ?https://pm.bsc.es/gitlab/DEEP-EST/apps/NBody repository and transfer it to a DEEP working directory.
The requirements of this application are shown in the following lists. The main requirements are:
The N-Body application has several versions which are built in different binaries. All of them divide the particle space into smaller blocks. MPI processes are divided into two groups: GPU processes and CPU processes. GPU processes are responsible for computing the forces between each pair of particles blocks, and then, these forces are sent to the CPU processes, where each process updates its particles blocks using the received forces. The particles and forces blocks are equally distributed amongst each MPI process in each group. Thus, each MPI process is in charge of computing the forces or updating the particles of a consecutive chunk of blocks.
This computation pattern is repeated during multiple timesteps. The communication pattern in each timestep consists in the fact that GPU processes exchange their particles with each other in a circular manner to compute the forces between their particles against the particles of all other GPU processes. For simplifying this pattern, the application uses a different MPI communicator for this circular exchanging. Once a GPU process finishes the computation of its forces, it sends the forces to the corresponding CPU processes, and then, it receives the updated particles. We perform separate sends/receives for each block.
The available versions are:
nbody.mpi.bin: Simple MPI parallel version using blocking MPI primitives for sending and
receiving each block of particles/forces.
nbody.mpi.ompss2.bin: Parallel version using MPI + OmpSs-2 tasks. Both computation
and communication phases are taskified. However, communication tasks (each one sending
or receiving a block) are serialized by an artificial dependency on a sentinel variable. This
is to prevent deadlocks between processes since communication tasks perform blocking MPI
calls.
nbody.mpi.ompss2.cuda.bin: The same as the previous version but offloading the tasks that
compute the forces between particles blocks to the available GPUs. The GPU processes offload those
computation tasks, which are the most compute-intensive parts of the program.
The calculate_forces_block_cuda task is annotated as a regular task (e.g., with their
dependencies) but implemented in CUDA. However, since it is Unified Memory, the user does
not need to move the data to/from the GPU device.
nbody.tampi.ompss2.bin: Parallel version using MPI + OmpSs-2 tasks + TAMPI library. This
version disables the artificial dependencies on the sentinel variable so that communication tasks can
run in parallel and overlap with computations. The TAMPI library is in charge of managing the blocking
MPI calls to avoid the blocking of the underlying execution resources.
nbody.tampi.ompss2.cuda.bin: A mix of the previous two variants where TAMPI is leveraged for
allowing the concurrent execution of communication tasks, and GPU processes offload the
compute-intensive tasks to the GPUs.
nbody.mpi.omp.bin: Parallel version using MPI + OpenMP tasks. Both computation and
communication phases are taskified. However, communication tasks (each one sending
or receiving a block) are serialized by an artificial dependency on a sentinel variable. This
is to prevent deadlocks between processes since communication tasks perform blocking MPI
calls.
nbody.mpi.omptarget.bin: The same as the previous version but offloading the tasks that
compute the forces between particles blocks to the available GPUs. The GPU processes offload those
computation tasks, which are the most compute-intensive parts of the program.
This is done through the omp target directive, declaring the corresponding dependencies, and
specifying the target as nowait (i.e., asynchronous offload). Additionally, the target directive
does not require the user to provide a CUDA implementation of the offloaded task. Finally, since we
are using the Unified Memory feature, we do not need to specify any data movement clause. We only
have to specify that the memory buffers are already device pointers (i.e., with is_device_ptr
clause). Note: This version is not compiled by default since it is still in a Work in
Progress state.
nbody.tampi.omp.bin: Parallel version using MPI + OpenMP tasks + TAMPI library. This
version disables the artificial dependencies on the sentinel variable so that communication tasks can
run in parallel and overlap computations. Since OpenMP only supports the non-blocking mechanism of
TAMPI, this version leverages non-blocking primitive calls. In this way, TAMPI library is in charge
of managing the non-blocking MPI operations to overlap communication and computation tasks
efficiently.
nbody.tampi.omptarget.bin: A mix of the previous two variants where TAMPI is leveraged for
allowing the concurrent execution of communication tasks, and GPU processes offload the
compute-intensive tasks to the GPUs. Note: This version is not compiled by default since it is
still in a Work in Progress state.
The simplest way to compile this application on the DEEP system is:
# Clone the benchmark's repository $ git clone https://pm.bsc.es/gitlab/DEEP-EST/apps/NBody.git $ cd NBody # Load the required environment (MPI, CUDA, OmpSs-2, OpenMP, etc.) # Needed only once per session $ source ./setenv_deep.sh # Compile all N-Body variants $ make
The benchmark versions are built with a specific block size, which is
decided at compilation time (i.e., the binary names contain the block
size). The default block size is 2048, but we can indicate
a different block size when compiling by doing:
$ make BS=1024
The next step is the execution of the benchmark on the DEEP system. Since this application targets the offloading of computation tasks to Unified Memory GPU devices, we must execute it in a DEEP partition that features this kind of devices. A good example is the dp-dam partition, where each node features:
In this case, we are going to request an interactive job in a dp-dam node.
All we need to is:
$ srun -p dp-dam -N 1 -n 8 -c 12 -t 01:00:00 --pty /bin/bash -i
With that command, we will be redirected to an interactive session in a dp-dam
node, exclusive for us. Furthermore, by indicating that configuration (i.e., -N,
-n and -c options), we are setting the default configuration for future srun
executions in that session. Thus, when executing an MPI binary via the srun command,
it is going to launch 8 processes and 12 CPUs per process by default. However, we
should be able to change the configuration (without overtaking the initial number
of resources) by overriding those parameters with new ones.
At this point, we are ready to execute the benchmark with multiple MPI processes.
The benchmark accepts several options. The most relevant options are the total
number of particles with -p, the number of timesteps with -t, and the
maximum number of GPU processes with -g. More options can be seen passing
the -h option. An example of execution is:
$ srun -n 8 -c 12 ./nbody.tampi.ompss2.cuda.2048bs.bin -t 100 -p 16384 -g 4
in which the application will perform 100 timesteps in 8 MPI processes with 12 cores per process (used by the OmpSs-2's runtime system). The maximum number of GPU processes is 4, so there will be 4 GPU processes and 4 CPU processes (all processes have access to GPU devices). Since the total number of particles is 16384 and the block size is 2048, each process will be in charge of computing/ updating 4096 forces/particles, which are 2 blocks.
In the CUDA variants, a process can belong to the GPU processes group if it has
access to at least one GPU device. However, in the case of the non-CUDA versions,
all processes can belong to the GPU processes group (i.e., we simulate the GPU
processes). For this reason, the application provides the -g option in order to
control the maximum number of GPU processes. By default, the number of GPU processes
will be half of the total number of processes. Also, note that the non-CUDA variants
cannot compute kernels on the GPU. In these cases, we have kept the structure of the
application, but we have replaced the CUDA tasks by regular CPU tasks, as a simulation.
Similarly, the OpenMP variants can be executed following the same steps but setting
the OMP_NUM_THREADS to the corresponding number of CPUs per process. As an example,
we could execute the following command:
$ OMP_NUM_THREADS=24 srun -n 4 -c 24 ./nbody.tampi.omp.2048bs.bin -t 100 -p 8912 -g 2
Finally, the submit.job script can be used to submit a non-interactive job into
the job scheduler system. Feel free to modify the script with other parameters or
job configurations. We can submit the script by doing:
$ sbatch submit.job