Changes between Version 18 and Version 19 of Public/User_Guide/Offloading_hybrid_apps

Timestamp:

Sep 17, 2019, 6:49:35 PM (5 years ago)

Author:

Kevin Sala

Comment:

—

Public/User_Guide/Offloading_hybrid_apps

@@ -13 +13 @@
 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.
+'''GPU devices'''.
 
 Moreover, some of the most modern GPU devices, such as NVIDIA Tesla V100,
-support the Unified Memory which facilitates the task of the users. With those
-devices, users do not have to move or copy the data to/from the GPU, and also,
-pointers at the host are the same at the device.
-
-For these reasons, parallel applications should try to take benefit from these
-GPU resources, and they should try to offload the most compute-intensive parts
-of the applications to the available GPUs. In this page, we are going to briefly
-explain the approaches proposed by the OpenMP and the !OmpSs-2 programming models
-to facilitate the offloading of computation tasks to the Unified Memory GPUs. Then,
-we show an hybrid MPI + OpenMP/OmpSs-2 benchmark that offloads some tasks, and that
-we can execute on the DEEP system.
-
-On the one hand, OpenMP provides the `target` directive which is the one used for
+support the '''Unified Memory''' which facilitates the task of the users. 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, parallel applications should take benefit from these GPU
+resources, and they should try to '''offload''' the most compute-intensive parts of
+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 an hybrid application with
+'''MPI+OpenMP''' and '''MPI+!OmpSs-2''' variants that offloads some of the computation
+tasks, and that can be executed on the DEEP system.
+
+On the one hand, '''OpenMP''' provides the `target` directive which is the one used for
 offloading computational parts of OpenMP programs to the GPUs. It provides multiple
-clauses for specifying 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
-inside using the `target` directive and without having to program that part 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 implementation in CUDA. 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 to the task is 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).
+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 inside 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 implementation in CUDA. 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 to 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).
 
 == N-Body Benchmark ==
@@ -205 +206 @@
 
 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 number
-of total 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 an execution is:
+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 an execution is:
 
 {{{#!bash
@@ -218 +219 @@
 GPU processes is 4, so there will be 4 CPU processes and 4 GPU processes (all
 processes have access to GPU devices). Since the total number of particles is 
-16384, each process will be in charge of computing/updating 4096 forces/particles,
-which are 2 blocks.
+16384 and the blocks 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
@@ -238 +239 @@
 }}}
 
-Finally, if you want to execute the benchmark without using an interactive session, you
-can modify the `submit.job` script and submit it into the job queue through the `sbatch`
-command.
+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. The script can be submitted by:
+
+{{{#!bash
+$ sbatch submit.job
+}}}
 
 == References ==