NVIDIA NVSHMEM 安装指南说明书-USB迷|专注于互联网分享

2024年7月22日发(作者：庆香岚)

NVIDIA NVSHMEM

Installation Guide

DG-08911-001_v2.0.3 | January 2021

Table of Contents

Chapter 1. 1

Chapter 2. Hardware And 2

2.1. 2

2.2. 2

Chapter 3. 3

3.1. .3

3.2. Building And 3

3.3. Using NVSHMEM In 4

3.3.1. Using NVSHMEM With Your C Or C++ 4

3.3.2. Using NVSHMEM With Your 5

3.4. Running .5

3.5. "Hello World" 6

Chapter 4. 8

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | ii

Chapter ew

NVIDIA

NVSHMEM

™

is a programming interface that implements a Partitioned Global

Address Space (PGAS) model across a cluster of NVIDIA GPUs. NVSHMEM provides an easy-

to-use interface to allocate memory that is symmetrically distributed across the GPUs. In

addition to a CPU-side interface, NVSHMEM also provides a CUDA kernel-side interface that

allows CUDA

threads to access any location in the symmetrically-distributed memory.

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 1

Chapter re And Software

Requirements

NVIDIA

NVSHMEM

™

has the following hardware and software requirements.

2.1. Hardware Requirements

NVSHMEM requires the following hardware:

‣

The

x86_64

ppc64leCPU

architecture.

‣

NVIDIA

Data Center GPU of NVIDIA Volta

™

GPU architecture or later.

Refer to /cuda-gpus for a complete list.

‣

All GPUs must be P2P-connected via NVLink/PCIe or via GPUDirect RDMA over InfiniBand/

RoCE with a Mellanox adapter (CX-4 or later).

Support for atomics requires a NVLink connection or a GPUDirect RDMA connection and

GDRCopy. See Software Requirements for more information.

2.2. Software Requirements

NVSHMEM requires the following software:

‣

64-bit Linux.

For a complete compatibility matrix, see the NVIDIA CUDA Installation Guide for Linux .

‣

CUDA 10.1 or later.

‣

Mellanox OFED.

‣

nv_peer_mem for GPUDirect RDMA

‣

PMI-1 (for example, Hydra), PMI-2 (for example, slurm), or a PMIx compatible launcher.

‣

(Optional)GDRCopy v2.0 or newer.

This software is required for atomics support on non-NVLink connections.

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 2

Chapter lation

3.1.

Procedure

Downloading NVSHMEM

To download NVSHMEM, go to NVSHMEM Downloads.

The extracted directory contains the following files and subdirectories:

File or Directory

src/

perftest/

examples/

scripts/

Description

Contains NVSHMEM sources and headers.

Contains tests showing use of NVSHMEM APIs

with performance reporting.

Contains examples showing use of some

common use cases of NVSHMEM.

Contains helper scripts, for example, script to

download, build and install Hydra.

NVSHMEM service level agreement (SLA).

3.2.

Procedure

Building And Installing NVSHMEM

the

CUDA_HOME

environment variable to point to the CUDA Toolkit.

the

GDRCOPY_HOME

environment variable to point to the GDRCopy installation.

To build without GDRCopy, set the environmental variable to

NVSHMEM_USE_GDRCOPY=0

Note: Without GDRCopy, atomics are only supported across NVLink connections.

MPI and/or SHMEM support is required, set

NVSHMEM_MPI_SUPPORT=1

and/or

NVSHMEM_SHMEM_SUPPORT=1

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 3

Installation

the

MPI_HOME

and

SHMEM_HOME

environment variables to point to the MPI and

OpenSHMEM installations, respectively.

NCCL will be used for host-initiated collectives, set

NVSHMEM_USE_NCCL=1

and

NCCL_HOME

to point to the NCCL installation.

You can use any NCCL 2.x version, and NVSHMEM hast been tested with NCCL 2.8.3-1.

the MPI library is neither OpenMPI nor its derivative, set

NVSHMEM_MPI_IS_OMPI=0

Note: Here is some additional information:

‣

When using OpenMPI and OSHMEM, the paths are the same.

‣

Other MPI and OpenSHMEM installations should also work.

‣

By default, MPI support is enabled, and OpenSHMEM support is disabled.

PMIx support is required, set

NVSHMEM_PMIX_SUPPORT=1

and

PMIX_HOME

to point to the

PMIx installation.

Note:

To use OSHMEM, OpenMPI needs to be built with UCX support.

NVSHMEM has been tested with OpenMPI 4.0.1 and UCX 1.8.0.

‣

PMI-1 and PMI-2 support is always included in the build, and PMI-1 is the default PMI.

To change the default

PMI NVSHMEM_DEFAULT_PMIX=1

(to select PMIx)

NVSHMEM_DEFAULT_PMI2=1

(to select PMI-2) can be set. At runtime,

NVSHMEM_BOOTSTRAP_PMI

can be used to override the default. The possible values are

PMIX, PMI-2, and PMI.

OpenMPI ships with its own copy of PMIx.

To avoid conflicting PMIx shared libraries, we recommend that you build a standalone

PMIx and configure OpenMPI with

--with-pmix=external

to point to that installation.

‣

specify the location where NVSHMEM will be installed, set

NVSHMEM_PREFIX

change the directory where NVSHMEM will be built, set

NVSHMEM_BUILDDIR

The default is

NVSHMEM_DIR/build

build and install the library, run

make -j8 install

3.3.

3.3.1.

Procedure

Using NVSHMEM In Your Applications

Using NVSHMEM With Your C Or C++

Program

nvshmem.h

and

nvshmemx.h

from

include/

to the

include/

and

lib/

paths.

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 4

Installation

M users: If your C or C++ program only uses NVSHMEM, install Hydra Process

Manager using the bash script

install_

under the

scripts/

directory. Provide

the download and install location as arguments, for example:

./install_

Use

nvshmrun

launcher under

bin/

(of the Hydra install path) to run the NVSHMEM job.

3.3.2. Using NVSHMEM With Your MPI Program

About this task

Note: The only MPI library currently tested is OpenMPI, however, derivatives of OpenMPI such

as SpectrumMPI as well as MPICH derivatives are expected to work.

To run a Hybrid MPI + NVSHMEM program, use the

mpirun

launcher available in the MPI

installation.

Similarly, NVSHMEM can be used from OpenSHMEM programs. In this case, you cannot use

the launchers that are in the NVSHMEM package. The only OpenSHMEM version that has

been tested is OSHMEM in OpenMPI. Other OpenSHMEM implementations such as Sandia

OpenSHMEM (SOS) should also work. To run the hybrid OpenSHMEM/NVSHMEM job, use

the

oshrun

launcher in the OpenMPI installation or follow the launcher specification of your

OpenSHMEM library.

3.4.

Procedure

Running Performance Tests

Before you can run performance tests, you first must build them.

the

CUDA_HOME

NVSHMEM_HOME

and

MPI_HOME

(if the NVSHMEM library was built with

NVSHMEM_MPI_SUPPORT=1

) environment variables to build NVSHMEM performance tests:

CUDA_HOME=

NVSHMEM_HOME=

MPI_HOME=

Configuring OpenMPI using the

-with-ucx

option is required for OpenMPI/OSHMEM

interoperability. If you have built NVSHMEM with MPI and OpenSHMEM support

(

NVSHMEM_MPI_SUPPORT=1

and

NVSHMEM_SHMEM_SUPPORT=1

), building

perftest/

also

requires MPI and OpenSHMEM support to be enabled.

Build without SHMEM interoperability: To build NVSHMEM performance tests without

SHMEM interoperability, set the environment variable

NVSHMEM_SHMEM_SUPPORT

. By

default, performance tests are installed under

perftest/perftest_install

. To install to

a different path, set

NVSHMEM_PERFTEST_INSTALL

to point to the desired path.

LD_LIBRARY_PATH

to point to

$CUDA_HOME/lib64

and

$MPI_HOME/lib

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 5

Installation

performance tests as NVSHMEM jobs (assuming Hydra is installed under

HYDRA_HOME

), hybrid MPI+NVSHMEM jobs, or hybrid OpenSHMEM+NVSHMEM jobs with the

following commands (using

perftest/device/pt-to-pt/

as an example):

NVSHMEM job using Hydra (PM-1):

$HYDRA_HOME/bin/nvshmrun -n

NIC accessible GPUs>

$NVSHMEM_PERFTEST_INSTALL/device/pt-to-pt/shmem_put_bw

NVSHMEM job using slurm:

srun -n

$NVSHMEM_PERFTEST_INSTALL/device/pt-to-pt/shmem_put_bw

Note: When slurm was built with a PMI that does not match the default of NVSHMEM, for

example, if slurm was built with PMIx support and

NVSHMEM_DEFAULT_PMIX=1

was not set

when building NVSHMEM,

NVSHMEM_BOOTSTRAP_PMI

can be used to override the default.

Possible values are PMIX, PMI-2, and PMI.

Hybrid MPI/NVSHMEM job:

$MPI_HOME/bin/mpirun -n

or InfiniBand NIC> -x NVSHMEMTEST_USE_MPI_LAUNCHER=1

$NVSHMEM_PERFTEST_INSTALL/device/pt-to-pt/shmem_put_bw

Hybrid OpenSHMEM/NVSHMEM job:

$MPI_HOME/bin/oshrun -n

or InfiniBand NIC> -x USE_SHMEM_IN_TEST=1

$NVSHMEM_PERFTEST_INSTALL/device/pt-to-pt/shmem_put_bw

3.5.

Procedure

"Hello World" Example

the following code as

#include

__global__ void simple_shift(int *destination) {

int mype = nvshmem_my_pe();

int npes = nvshmem_n_pes();

int peer = (mype + 1) % npes;

nvshmem_int_p(destination, mype, peer);

}

int main(void) {

int mype_node, msg;

cudaStream_t stream;

nvshmem_init();

mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);

cudaSetDevice(mype_node);

cudaStreamCreate(&stream);

int *destination = (int *) nvshmem_malloc(sizeof(int));

simple_shift<<<1, 1, 0, stream>>>(destination);

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 6

Installation

nvshmemx_barrier_all_on_stream(stream);

cudaMemcpyAsync(&msg, destination, sizeof(int), cudaMemcpyDeviceToHost,

stream);

cudaStreamSynchronize(stream);

printf("%d: received message %dn", nvshmem_my_pe(), msg);

nvshmem_free(destination);

nvshmem_finalize();

return 0;

}

with the following command:

nvcc -rdc=true -ccbin g++ -gencode=$NVCC_GENCODE -I $NVSHMEM_HOME/include

-o -L $NVSHMEM_HOME/lib -lnvshmem

-lcuda

Where

arch=compute_70,code=sm_70

is the value of

NVCC_GENCODE

for V100 GPUs.

the

nvshmemHelloWorld

sample with one of the following commands:

When running on a single host with 2 GPUs (connected by PCI-E, NVLink or Infiniband):

$HYDRA_HOME/bin/nvshmrun -n 2 -ppn 2 ./

When running on two hosts with 1 GPU per host (connected by InfiniBand):

$HYDRA_HOME/bin/nvshmrun -n 2 -ppn 1 –-hosts hostname1,hostname2 ./

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 7

Chapter t

Report bugs and submit feature requests using NVONLINE or by emailing

******************.

NVIDIA NVSHMEMDG-08911-001_v2.0.3 | 8

Notice

This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA

Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this

document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any

infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material

(defined below), code, or functionality.

This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA

Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this

document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any

infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material

(defined below), code, or functionality.

Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.

NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed

in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any

customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed

either directly or indirectly by this document.

NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications

where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA

accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.

NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product

is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document,

ensure the product is suitable and fit for the application planned by customer, and perform the necessary testing for the application in order to avoid a default of

the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional

or different conditions and/or requirements beyond those contained in this document. NVIDIA accepts no liability related to any default, damage, costs, or problem

which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document or (ii) customer product designs.

VESA DisplayPort

DisplayPort and DisplayPort Compliance Logo, DisplayPort Compliance Logo for Dual-mode Sources, and DisplayPort Compliance Logo for Active Cables are

trademarks owned by the Video Electronics Standards Association in the United States and other countries.

HDMI

HDMI, the HDMI logo, and High-Definition Multimedia Interface are trademarks or registered trademarks of HDMI Licensing LLC.

OpenCL

OpenCL is a trademark of Apple Inc. used under license to the Khronos Group Inc.

NVIDIA Corporation | 2788 San Tomas Expressway, Santa Clara, CA 95051

Trademarks

NVIDIA, the NVIDIA logo, and CUDA, CUDA Toolkit, GPU, Kepler, Mellanox, NVLink, NVSHMEM, and Tesla are trademarks and/or registered trademarks of NVIDIA

Corporation in the United States and other countries. Other company and product names may be trademarks of the respective companies with which they are

associated.