Skip to content
This repository has been archived by the owner on Dec 18, 2020. It is now read-only.

ospray/module_mpi

Repository files navigation

OSPRay MPI Module

This is release v2.4.0 of Intel® OSPRay's MPI Module. For changes and new features see the changelog. Visit http://www.ospray.org for more information.

OSPRay MPI Module Overview

Intel OSPRay is an open source, scalable, and portable ray tracing engine for high-performance, high-fidelity visualization on Intel Architecture CPUs. OSPRay is part of the Intel oneAPI Rendering Toolkit and is released under the permissive Apache 2.0 license.

The purpose of the MPI module for OSPRay is to provide distributed rendering capabilities for OSPRay. The module enables image- and data-parallel rendering across HPC clusters using MPI, allowing applications to transparently distribute rendering work, or to render data sets which are too large to fit in memory on a single machine.

OSPRay internally builds on top of [Intel Embree]((https://www.embree.org/) and ISPC (Intel SPMD Program Compiler), and fully exploits modern instruction sets like Intel SSE4, AVX, AVX2, and AVX-512 to achieve high rendering performance, thus a CPU with support for at least SSE4.1 is required to run OSPRay. The MPI module requires an MPI library which supports at least MPI_THREAD_SERIALIZED.

OSPRay Support and Contact

OSPRay is under active development, and though we do our best to guarantee stable release versions a certain number of bugs, as-yet-missing features, inconsistencies, or any other issues are still possible. Should you find any such issues please report them immediately via OSPRay's GitHub Issue Tracker (or, if you should happen to have a fix for it, you can also send us a pull request). For bugs specific to the MPI module and distributed rendering, please report them to the MPI Module's GitHub Issue Tracker For missing features please contact us via email at [email protected].

To receive release announcements simply “Watch” the OSPRay MPI module repository on GitHub.

Join the chat at https://gitter.im/ospray/ospray

Building the MPI Module

The latest MPI module sources are always available at the MPI Module GitHub repository. The default master branch should always point to the latest bugfix release.

Prerequisites

OSPRay v2.2.x is required, and can be built from source following the instructions on the OSPRay GitHub Repository. The MPI module can also be built together with OSPRay using the superbuild by setting BUILD_OSPRAY_MODULE_MPI=ON. After building OSPRay, you can build this module by pointing it to your OSPRay, OpenVKL and rkcommon install directories. Applications using the MPI module can then be run by ensuring the module is in your library path and loading the module via ospLoadModule or using the command line options below. When building the module with BUILD_TUTORIALS=ON the MPI module's tutorial apps will be built, which demonstrate the various features of the module.

Documentation

The MPI module provides two OSPRay devices to allow applications to leverage distributed rendering capabilities. The mpiOffload device provides transparent image-parallel rendering, where the same OSPRay application written for local rendering can be replicated across multiple nodes to distribute the rendering work. The mpiDistributed device allows MPI distributed applications to use OSPRay for distributed rendering, where each rank can render and independent piece of a global data set, or hybrid rendering where ranks partially or completely share data.

MPI Offload Rendering

The mpiOffload device can be used to distribute image rendering tasks across a cluster without requiring modifications to the application itself. Existing applications using OSPRay for local rendering simply be passed command line arguments to load the module and indicate that the mpiOffload device should be used for image-parallel rendering. To load the module, pass --osp:load-modules=mpi, to select the MPIOffloadDevice, pass --osp:device=mpiOffload. For example, the ospExamples application can be run as:

mpirun -n <N> ./ospExamples --osp:load-modules=mpi --osp:device=mpiOffload

and will automatically distribute the image rendering tasks among the corresponding N nodes. Note that in this configuration rank 0 will act as a master/application rank, and will run the user application code but not perform rendering locally. Thus, a minimum of 2 ranks are required, one master to run the application and one worker to perform the rendering. Running with 3 ranks for example would now distribute half the image rendering work to rank 1 and half to rank 2.

If more control is required over the placement of ranks to nodes, or you want to run a worker rank on the master node as well you can run the application and the ospray_mpi_worker program through MPI's MPMD mode. The ospray_mpi_worker will load the MPI module and select the offload device by default.

mpirun -n 1 ./ospExamples --osp:load-modules=mpi --osp:device=mpiOffload \
  : -n <N> ./ospray_mpi_worker

Finally, you can also run the workers in a server mode on a remote machine and connect your application to them over a socket. This allows remote rendering on a large cluster while displaying on a local machine (e.g., a laptop) where the two devices may not be able to connect over MPI. First, launch the workers in mpi-listen mode:

mpirun -n <N> ./ospray_mpi_worker --osp:device-params=mpiMode:mpi-listen

The workers will print out a port number to connect to, e.g., #osp: Listening on port ##### You can then run your application in the mpi-connect mode, and pass the host name of the first worker rank and this port number to the device:

./ospExamples --osp:load-modules=mpi --osp:device=mpiOffload \
  --osp:device-params=mpiMode:mpi-connect,host:<worker rank 0 host>,port:<port printed above>

If initializing the mpiOffload device manually, or passing parameters through the command line, the following parameters can be set:

Type Name Default Description
string mpiMode mpi The mode to communicate with the worker ranks. mpi will assume you're launching the application and workers in the same mpi command (or split launch command). mpi-listen can be passed to the workers, indicating they should wait and listen for a connection from the application. mpi-connect can be passed to the application, indicating it should connect to the first worker at host and port to connect to the workers
string host none, optional On the app rank, specify the host worker 0 is on to connect to in mpi-connect mode
int port none, optional On the app rank, specify the port worker 0 is listening on to connect in mpi-connect mode
uint maxCommandBufferEntries 8192 Set the max number of commands to buffer before submitting the command buffer to the workers
uint commandBufferSize 512MiB Set the max command buffer size to allow. Units are in MiB. Max size is 1.8GiB
uint maxInlineDataSize 32MiB Set the max size of an OSPData which can be inline'd into the command buffer instead of being sent separately. Max size is half the commandBufferSize. Units are in MiB

: Parameters specific to the mpiOffload Device.

The maxCommandBufferEntries, commandBufferSize, and maxInlineDataSize can also be set via the environment variables: OSPRAY_MPI_MAX_COMMAND_BUFFER_ENTRIES, OSPRAY_MPI_COMMAND_BUFFER_SIZE, and OSPRAY_MPI_MAX_INLINE_DATA_SIZE, respectively.

MPI Distributed Rendering

While MPI Offload rendering is used to transparently distribute rendering work without requiring modification to the application, MPI Distributed rendering is targetted at use of OSPRay within MPI-parallel applications. The MPI distributed device can be selected by loading the mpi module, and manually creating and using an instance of the mpiDistributed device.

ospLoadModule("mpi");

OSPDevice mpiDevice = ospNewDevice("mpiDistributed");
ospDeviceCommit(mpiDevice);
ospSetCurrentDevice(mpiDevice);

Your application can either initialize MPI before-hand, ensuring that MPI_THREAD_SERIALIZED or higher is supported, or allow the device to initialize MPI on commit. Thread multiple support is required if your application will make MPI calls while rendering asynchronously with OSPRay. When using the distributed device each rank can specify independent local data using the OSPRay API, as if rendering locally. However, when calling ospRenderFrameAsync the ranks will work collectively to render the data. The distributed device supports both image-parallel, where the data is replicated, and data-parallel, where the data is distributed, rendering modes. The mpiDistributed device will by default use each rank in MPI_COMM_WORLD as a render worker; however, it can also take a specific MPI communicator to use as the world communicator. Only those ranks in the specified communicator will participate in rendering.

Type Name Default Description
void* worldCommunicator MPI_COMM_WORLD The MPI communicator which the OSPRay workers should treat as their world

: Parameters specific to the distributed mpiDistributed Device.

Type Name Default Description
OSPBox3f[] region NULL A list of bounding boxes which bound the owned local data to be rendered by the rank

: Parameters specific to the distributed OSPWorld.

Type Name Default Description
aoSamples int 0 The number of AO samples to take per-pixel
aoRadius float 1e20f The AO ray length to use. Note that if the AO ray would have crossed a rank boundary and ghost geometry is not available, there will be visible artifacts in the shading.

: Parameters specific to the mpiRaycast renderer.

Image Parallel Rendering in the MPI Distributed Device

If all ranks specify exactly the same data, the distributed device can be used for image-parallel rendering. This works identical to the offload device, except that the MPI-aware application is able to load data in parallel on each rank rather than loading on the master and shipping data out to the workers. When a parallel file system is available, this can improve data load times. Image-parallel rendering is selected by specifying the same data on each rank, and using any of the existing local renderers (e.g., scivis, pathtracer). See ospMPIDistributedTutorialReplicatedData for an example.

Data Parallel Rendering in the MPI Distributed Device

The MPI Distributed device also supports data-parallel rendering with sort-last compositing. Each rank can specify a different piece of data, as long as the bounding boxes of each rank's data are non-overlapping. The rest of the scene setup is similar to local rendering; however, for distributed rendering only the mpiRaycast renderer is supported. This renderer implements a subset of the scivis rendering features which are suitable for implementation in a distributed environment.

By default the aggregate bounding box of the instances in the local world will be used as the bounds of that rank's data. However, when using ghost zones for volume interpolation, geometry or ambient occlusion, each rank's data can overlap. To clip these non-owned overlap regions out a set of regions (the region parameter) can pass as a parameter to the OSPWorld being rendered. Each rank can specify one or more non-overlapping box3f's which bound the portions of its local data which it is reponsible for rendering. See the ospMPIDistributedTutorialStructuredVolume for an example.

Finally, the MPI distributed device also supports hybrid-parallel rendering, where multiple ranks can share a single piece of data. For each shared piece of data the rendering work will be assigned image-parallel among the ranks. Partially-shared regions are determined by finding those ranks specifying data with the same bounds (matching regions) and merging them. See the ospMPIDistributedTutorialPartiallyReplicatedData for an example.

Interaction With User Modules

The MPI Offload rendering mode trivially supports user modules, with the caveat that attempting to share data directly with the application (e.g., passing a void* or other tricks to the module) will not work in a distributed environment. Instead, use the ospNewSharedData API to share data from the application with OSPRay, which will in turn be copied over the network to the workers.

The MPI Distributed device also supports user modules, as all that is required for compositing the distributed data are the bounds of each rank's local data.