xCAT/xcat-core
2
0
Fork 0
mirror of https://github.com/xcat2/xcat-core.git synced 2025-05-22 03:32:04 +00:00
Code Issues Releases Wiki Activity

Refactor the GPU CUDA documentation. Organized for both

Browse Source 
RHEL and Ubuntu sharing common pages.  Made sure the documentation
reflects the chanages made for diskless support in Issue #314
and Issue #316
This commit is contained in:
Victor Hu 2015-10-28 16:38:49 -04:00
parent 0d023445d0
commit 6b54a9dd76
14 changed files with 639 additions and 646 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

307
docs/source/advanced/gpu/cuda_rhel.rst
View File

@ -1,307 +0,0 @@
CUDA Installation
=================
Overview
--------
CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model created by NVIDIA. It can be used by the graphics processing units (GPUs) for general purpose processing.
xCAT supports CUDA installation for Ubuntu and rhel7.2 on PowerNV (p8le node with NVIDIA GPU Support) for both diskless and diskful nodes. The CUDA packages provided by NVIDIA include both the runtime libraries for computing and development tools for programming and monitoring. The full package set is very large, so in xCAT, it's suggested that the packages be split into two package sets:
#. **cudaruntime** package set
#. **cudafull** package set
It's suggested to only installing the **cudaruntime** package set on the Compute Nodes (CNs), and the **cudafull** package set on the Management Node or the monitor/development nodes.
In this documentation, xCAT will provide CUDA installation based on rhel7.2 Power 8 firestone nodes. User can find Ubuntu CUDA installation document here: http://sourceforge.net/p/xcat/wiki/xCAT_P8LE_cuda_installing/
CUDA Repository
---------------
The NIVIDIA CUDA Toolkit is available at http://developer.nvidia.com/cuda-downloads. User can download the CUDA Toolkit based on the target platform.
Prepare a local repository directory which contains all the CUDA packages and repository meta-data. User can either
* Create a repository on the Management node installing the CUDA toolkit: ::
mkdir -p /install/cuda-repo/cuda-7-5
rpm -ivh cuda-repo-rhel7-7-5-local-7.5-18.ppc64le.rpm
cp /var/cuda-repo-7-5-local/* /install/cuda-repo/cuda-7-5
cd /install/cuda-repo/cuda-7-5
createrepo .
* Create a repository on the Management node without installing the CUDA toolkit: ::
mkdir -p /tmp/cuda
cd /tmp/cuda
rpm2cpio /root/cuda-repo-rhel7-7-5-local-7.5-18.ppc64le.rpm | cpio -i -d
cp /tmp/cuda/var/cuda-repo-7-5-local/* /install/cuda-repo/cuda-7-5
cd /install/cuda-repo/cuda-7-5
createrepo .
The NVIDIA driver RPM packages depend on other external packages, such as DKMS and maybe EPEL (firestone node doesn't need this package). Users need to download those package to the directory. ::
mkdir -p /install/cuda-repo/cuda-deps
ls -ltr /install/cuda-repo/cuda-deps
-rw-r--r-- 1 root root 79048 Oct 5 10:58 dkms-2.2.0.3-30.git.7c3e7c5.el7.noarch.rpm
cd /install/cuda-repo/cuda-deps
createrepo .
CUDA osimage
------------
User can generate a new CUDA osimage object based on another osimage definition or just modify an existing osimage. xCAT provides some sample CUDA pkglist files:
* diskful provisioning in ``/opt/xcat/share/install/rh/`` for ``cudafull`` and ``cudaruntime``: ::
#cat /opt/xcat/share/xcat/install/rh/cudafull.rhels7.pkglist
#INCLUDE:compute.rhels7.pkglist#
#For Cuda 7.5
kernel-devel
gcc
pciutils
dkms
cuda
#cat /opt/xcat/share/xcat/install/rh/cudaruntime.rhels7.pkglist
#INCLUDE:compute.rhels7.pkglist#
#For Cuda 7.5
kernel-devel
gcc
pciutils
dkms
cuda-runtime-7-5
* diskless provisioning in ``/opt/xcat/share/xcat/netboot/rh`` for ``cudafull`` and ``cudaruntime``: ::
#cat /opt/xcat/share/xcat/netboot/rh/cudafull.rhels7.ppc64le.pkglist
#INCLUDE:compute.rhels7.ppc64.pkglist#
#For Cuda 7.5
kernel-devel
gcc
pciutils
dkms
cuda
#cat /opt/xcat/share/xcat/netboot/rh/cudaruntime.rhels7.ppc64le.pkglist
#INCLUDE:compute.rhels7.ppc64.pkglist#
#For Cuda 7.5
kernel-devel
gcc
pciutils
dkms
cuda-runtime-7-5
**NOTE: After CUDA are installed, the nodes require a reboot**
* For the diskful installation, the CUDA packages should be included in the ``pkglist`` field so a reboot happens automatically after the OS is installed.
* For the diskless installation, the CUDA package can be included either in ``otherpkglist`` or ``pktlist`` fields.
The following are some sample osimage definitions:
* The diskful cudafull installation osimage object. ::
#lsdef -t osimage rhels7.2-ppc64le-install-cudafull
Object name: rhels7.2-ppc64le-install-cudafull
imagetype=linux
osarch=ppc64le
osdistroname=rhels7.2-ppc64le
osname=Linux
osvers=rhels7.2
otherpkgdir=/install/post/otherpkgs/rhels7.2/ppc64le
pkgdir=/install/rhels7.2/ppc64le,/install/cuda-repo
pkglist=/opt/xcat/share/xcat/install/rh/cudafull.rhels7.pkglist
profile=compute
provmethod=install
template=/opt/xcat/share/xcat/install/rh/compute.rhels7.tmpl
* The diskful cudaruntime installation osimage object. ::
#lsdef -t osimage rhels7.2-ppc64le-install-cudaruntime
Object name: rhels7.2-ppc64le-install-cudaruntime
imagetype=linux
osarch=ppc64le
osdistroname=rhels7.2-ppc64le
osname=Linux
osvers=rhels7.2
otherpkgdir=/install/post/otherpkgs/rhels7.2/ppc64le
pkgdir=/install/rhels7.2/ppc64le,/install/cuda-repo
pkglist=/opt/xcat/share/xcat/install/rh/cudairuntime.rhels7.pkglist
profile=compute
provmethod=install
template=/opt/xcat/share/xcat/install/rh/compute.rhels7.tmpl
* The diskless cudafull installation osimage object. ::
#lsdef -t osimage rhels7.2-ppc64le-netboot-cudafull
Object name: rhels7.2-ppc64le-netboot-cudafull
imagetype=linux
osarch=ppc64le
osdistroname=rhels7.2-ppc64le
osname=Linux
osvers=rhels7.2
otherpkgdir=/install/post/otherpkgs/rhels7.2/ppc64le
permission=755
pkgdir=/install/rhels7.2/ppc64le,/install/cuda-repo
pkglist=/opt/xcat/share/xcat/netboot/rh/cudafull.rhels7.ppc64le.pkglist
postinstall=/opt/xcat/share/xcat/netboot/rh/compute.rhels7.ppc64le.postinstall
profile=compute
provmethod=netboot
rootimgdir=/install/netboot/rhels7.2/ppc64le/compute
* The diskless cudaruntime installation osimage object. ::
#lsdef -t osimage rhels7.2-ppc64le-netboot-cudaruntime
Object name: rhels7.2-ppc64le-netboot-cudaruntime
imagetype=linux
osarch=ppc64le
osdistroname=rhels7.2-ppc64le
osname=Linux
osvers=rhels7.2
otherpkgdir=/install/post/otherpkgs/rhels7.2/ppc64le
permission=755
pkgdir=/install/rhels7.2/ppc64le,/install/cuda-repo
pkglist=/opt/xcat/share/xcat/netboot/rh/cudaruntime.rhels7.ppc64le.pkglist
postinstall=/opt/xcat/share/xcat/netboot/rh/compute.rhels7.ppc64le.postinstall
profile=compute
provmethod=netboot
rootimgdir=/install/netboot/rhels7.2/ppc64le/compute
Deployment of CUDA node
-----------------------
* To provision diskful nodes: ::
nodeset <node> osimage=rhels7.2-ppc64le-install-cudafull
rsetboot <node> net
rpower <node> boot
* To provision diskless nodes: ::
genimage rhels7.2-ppc64le-netboot-cudafull
packimage rhels7.2-ppc64le-netboot-cudafull
nodeset <node> osimage=rhels7.2-ppc64le-netboot-cudafull
rsetboot <node> net
rpower <node> boot
Verification of CUDA Installation
---------------------------------
**NOTE** For ``cudaruntime`` installation, it only provide the basic libraries that can bee used by other applications which works with GPU. The following verification will not apply to ``cudaruntime`` installations.
After compute node booted, The Environment variable has to be set in order to use the CUDA toolkits. The PATH variable needs to include ``/usr/local/cuda-7.5/bin`` and LD_LIBRARY_PATH variable needs to contain ``/usr/local/cuda-7.5/lib64`` on a 64-bit system, and ``/usr/local/cuda-7.5`` on a 32-bit system.
* To change the environment variables for 64-bit operating systems ::
export PATH=/usr/local/cuda-7.5/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH
* To change the environment variable for 32-bit operating systems ::
export PATH=/usr/local/cuda-7.5/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib:$LD_LIBRARY_PATH
After Environment variables are set correctly, user can verify the CUDA installation by
* Verify the Driver Version ::
#cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX ppc64le Kernel Module 352.39 Fri Aug 14 17:10:41 PDT 2015
GCC version: gcc version 4.8.5 20150623 (Red Hat 4.8.5-4) (GCC)
* The version of the CUDA Toolkits ::
#nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Tue_Aug_11_14:31:50_CDT_2015
Cuda compilation tools, release 7.5, V7.5.17
* Compiling the Examples, then can run a `deviceQuery` or `bandwidthTest` or other commands under the bin directory to ensure the system and the CUDA-capable device are able to communicate correctly ::
# mkdir -p /tmp/cuda
# cuda-install-samples-7.5.sh /tmp/cuda
# cd /tmp/cuda/NVIDIA_CUDA-7.5_Samples
# make
# cd bin/ppc64le/linux/release
# ./deviceQuery
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 4 CUDA Capable device(s)
Device 0: "Tesla K80"
CUDA Driver Version / Runtime Version 7.5 / 7.5
CUDA Capability Major/Minor version number: 3.7
Total amount of global memory: 11520 MBytes (12079136768 bytes)
(13) Multiprocessors, (192) CUDA Cores/MP: 2496 CUDA Cores
GPU Max Clock rate: 824 MHz (0.82 GHz)
Memory Clock rate: 2505 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 1572864 bytes
............
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 7.5, CUDA Runtime Version = 7.5, NumDevs = 4, Device0 = Tesla K80, Device1 = Tesla K80, Device2 = Tesla K80, Device3 = Tesla K80
Result = PASS
# ./bandwidthTest
[CUDA Bandwidth Test] - Starting...
Running on...
Device 0: Tesla K80
Quick Mode
Host to Device Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 7765.1
Device to Host Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 7759.6
Device to Device Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 141485.3
Result = PASS
NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.
* The tool `nvidia-smi` providied by NVIDIA driver can be used to do GPU management and monitoring. ::
#nvidia-smi -q
==============NVSMI LOG==============
Timestamp : Mon Oct 5 13:43:39 2015
Driver Version : 352.39
Attached GPUs : 4
GPU 0000:03:00.0
Product Name : Tesla K80
Product Brand : Tesla
...........................

337
docs/source/advanced/gpu/cuda_ubuntu.rst
View File

@ -1,337 +0,0 @@
CUDA Installation Based on Ubuntu
=================================
Overview
--------
CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model created by NVIDIA. It can be used by the graphics processing units (GPUs) for general purpose processing.
xCAT supports CUDA installation for Ubuntu 14.04.3 and RHEL 7.2LE on PowerNV (Non-Virtualized) for both diskless and diskful nodes. The CUDA packages provided by NVIDIA include both the runtime libraries for computing and development tools for programming and monitoring. The full package set is very large, so in xCAT, it's suggested that the packages be split into two package sets:
#. **cudaruntime** package set
#. **cudafull** package set
It's suggested to only installing the **cudaruntime** package set on the Compute Nodes (CNs), and the **cudafull** package set on the Management Node or the monitor/development nodes.
In this documentation, xCAT will provide CUDA installation based on Ubuntu 14.04.3 running on IBM Power Systems S822LC nodes.
CUDA Repository
---------------
Currently, there are 2 types of Ubuntu repos for installing cuda-7-0 on p8LE hardware: `The online repo <http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/ppc64el/cuda-repo-ubuntu1404_7.0-28_ppc64el.deb>`_ and `The local package repo <http://developer.download.nvidia.com/compute/cuda/7_0/Prod/local_installers/rpmdeb/cuda-repo-ubuntu1404-7-0-local_7.0-28_ppc64el.deb>`_.
**The online repo**
The online repo will provide a sourcelist entry which includes the URL with the location of the cuda packages. The online repo can be used directly by Compute Nodes. The source.list entry will be similar to: ::
deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1410/ppc64el /
**The local package repo**
A local package repo will contain all of the cuda packages.The admin can either simply install the local repo (need to copy the whole /var/cuda-repo-7-0-local/ to /install/cuda-repo/)or extract the cuda packages into the local repo with the following command: ::
dpkg -x cuda-repo-ubuntu14xx-7-0-local_7.0-28_ppc64el.deb /install/cuda-repo/
The following repos will be used in the test environment:
* "/install/ubuntu14.04.2/ppc64el": The OS image package directory
* "http://ports.ubuntu.com/ubuntu-ports": The internet mirror, if there is local mirror available, it can be replaced
* "http://10.3.5.10/install/cuda-repo/var/cuda-repo-7-0-local /": The repo for cuda, you can replaced it with online cuda repo.
CUDA osimage
------------
User can generate a new CUDA osimage object based on another osimage definition or just modify an existing osimage. xCAT provides some sample CUDA pkglist files:
* diskful provisioning in ``/opt/xcat/share/install/ubuntu/`` for ``cudafull`` and ``cudaruntime``: ::
#cat /opt/xcat/share/xcat/install/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist
#INCLUDE:compute.ubuntu14.04.3.ppc64el.pkglist#
linux-headers-generic-lts-utopic
build-essential
dkms
zlib1g-dev
cuda
#cat /opt/xcat/share/xcat/install/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist
#INCLUDE:compute.ubuntu14.04.3.ppc64el.pkglist#
linux-headers-generic-lts-utopic
build-essential
dkms
zlib1g-dev
cuda-runtime-7-0
* diskless provisioning in ``/opt/xcat/share/xcat/netboot/ubuntu`` for ``cudafull`` and ``cudaruntime``: ::
#cat /opt/xcat/share/xcat/netboot/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist
#INCLUDE:compute.ubuntu14.04.3.ppc64el.pkglist#
linux-headers-generic-lts-utopic
build-essential
zlib1g-dev
dkms
#cat /opt/xcat/share/xcat/netboot/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist
#INCLUDE:compute.ubuntu14.04.3.ppc64el.pkglist#
linux-headers-generic-lts-utopic
build-essential
zlib1g-dev
dkms
The following are some sample osimage definitions:
* The diskful cudafull installation osimage object. ::
#lsdef -t osimage ubuntu14.04.3-ppc64el-install-cudafull
Object name: ubuntu14.04.3-ppc64el-install-cudafull
imagetype=linux
osarch=ppc64el
osname=Linux
osvers=ubuntu14.04.3
otherpkgdir=/install/post/otherpkgs/ubuntu14.04.3/ppc64el
pkgdir=http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main,http://10.3.5.10/install/cuda-repo/var/cuda-repo-7-0-local /,/install/ubuntu14.04.3/ppc64el
pkglist=/opt/xcat/share/xcat/install/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist
profile=cudafull
provmethod=install
template=/opt/xcat/share/xcat/install/ubuntu/cudafull.tmpl
* The diskful cudaruntime installation osimage object. ::
#lsdef -t osimage ubuntu14.04.3-ppc64el-install-cudaruntime
Object name: ubuntu14.04.3-ppc64el-install-cudaruntime
imagetype=linux
osarch=ppc64el
osname=Linux
osvers=ubuntu14.04.3
otherpkgdir=/install/post/otherpkgs/ubuntu14.04.3/ppc64el
pkgdir=http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main,http://10.3.5.10/install/cuda-repo/var/cuda-repo-7-0-local /,/install/ubuntu14.04.3/ppc64el
pkglist=/opt/xcat/share/xcat/install/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist
profile=cudaruntime
provmethod=install
template=/opt/xcat/share/xcat/install/ubuntu/cudaruntime.tmpl
* The diskless cudafull installation osimage object. ::
#Object name: ubuntu14.04.3-ppc64el-netboot-cudafull
imagetype=linux
osarch=ppc64el
osname=Linux
osvers=ubuntu14.04.3
otherpkgdir=http://10.3.5.10/install/cuda-repo/var/cuda-repo-7-0-local /
otherpkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudafull.otherpkgs.pkglist
permission=755
pkgdir=http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main,/install/ubuntu14.04.3/ppc64el
pkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist
profile=cudafull
provmethod=netboot
rootimgdir=/install/netboot/ubuntu14.04.3/ppc64el/cudafull
* The diskless cudaruntime installation osimage object. ::
#Object name: ubuntu14.04.3-ppc64el-netboot-cudaruntime
imagetype=linux
osarch=ppc64el
osname=Linux
osvers=ubuntu14.04.3
otherpkgdir=http://10.3.5.10/install/cuda-repo/var/cuda-repo-7-0-local /
otherpkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudaruntime.otherpkgs.pkglist
permission=755
pkgdir=http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main,/install/ubuntu14.04.3/ppc64el
pkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist
profile=cudaruntime
provmethod=netboot
rootimgdir=/install/netboot/ubuntu14.04.3/ppc64el/cudaruntime
**Use addcudakey postscript to install GPGKEY for cuda packages**
In order to access the cuda repo and authorize it, you will need to import the cuda GPGKEY into the apt key trust list.The following command can be used to add a postscript for a node that will install cuda. ::
chdef <node> -p postscripts=addcudakey
**Install NVML (optional, for nodes which need to compile cuda related applications)**
The NVIDIA Management Library (NVML) is a C-based programmatic interface for monitoring and managing various states within NVIDIA Tesla™ GPUs. It is intended to be a platform for building 3rd party applications.
The NVML can be download from http://developer.download.nvidia.com/compute/cuda/7_0/Prod/local_installers/cuda_346.46_gdk_linux.run.
After download NVML and put it under /install/postscripts on MN, the following steps can be used to have NVML installed after the node is installed and rebooted if needed. ::
chmod +x /install/postscripts/cuda_346.46_gdk_linux.run
chdef <node> -p postbootscripts="cuda_346.46_gdk_linux.run --silent --installdir=<you_desired_dir>"
Deployment of CUDA node
-----------------------
* To provision diskful nodes: ::
nodeset <node> osimage=<diskfull_osimage_object_name>
rsetboot <node> net
rpower <node> boot
* To provision diskless nodes:
To generate stateless image for a diskless installation, the acpid is needed to be installed on MN or the host on which you generate stateless image. ::
apt-get install -y acpid
Then, use the following commands to generate stateless image and pack it. ::
genimage <diskless_osimage_object_name>
packimage <diskless_osimage_object_name>
nodeset <node> osimage=<diskless_osimage_object_name>
rsetboot <node> net
rpower <node> boot
Verification of CUDA Installation
---------------------------------
The command below can be used to display GPU or Unit info on the node. ::
nvidia-smi -q
Verify the Driver Version. ::
# cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX ppc64le Kernel Module 346.46 Tue Feb 17 17:18:33 PST 2015
GCC version: gcc version 4.8.4 (Ubuntu 4.8.4-2ubuntu1~14.04)
**GPU management and monitoring**
The tool "nvidia-smi" provided by NVIDIA driver can be used to do GPU management and monitoring, but it can only be run on the host where GPU hardware, CUDA and NVIDIA driver is installed. The xdsh can be used to run "nvidia-smi" on GPU host remotely from xCAT management node.
The using of xdsh will be like this: ::
# xdsh p8le-42l "nvidia-smi -i 0 --query-gpu=name,serial,uuid --format=csv,noheader"
p8le-42l: Tesla K40m, 0324114102927, GPU-8750df00-40e1-8a39-9fd8-9c29905fa127
Some of the useful nvidia-smi command for monitoring and managing of GPU are as belows, for more information, pls read nvidia-smi manpage.
* For monitoring: ::
*The number of NVIDIA GPUs in the system.
nvidia-smi --query-gpu=count --format=csv,noheader
*The version of the installed NVIDIA display driver
nvidia-smi -i 0 --query-gpu=driver_version --format=csv,noheader
*The BIOS of the GPU board
nvidia-smi -i 0 --query-gpu=vbios_version --format=csv,noheader
*Product name, serial number and UUID of the GPU
nvidia-smi -i 0 --query-gpu=name,serial,uuid --format=csv,noheader
*Fan speed
nvidia-smi -i 0 --query-gpu=fan.speed --format=csv,noheader
*The compute mode flag indicates whether individual or multiple compute applications may run on the GPU. Also known as exclusivity modes
nvidia-smi -i 0 --query-gpu=compute_mode --format=csv,noheader
*Percent of time over the past sample period during which one or more kernels was executing on the GPU
nvidia-smi -i 0 --query-gpu=utilization.gpu --format=csv,noheader
*Total errors detected across entire chip. Sum of device_memory, register_file, l1_cache, l2_cache and texture_memory
nvidia-smi -i 0 --query-gpu=ecc.errors.corrected.aggregate.total --format=csv,noheader
*Core GPU temperature, in degrees C
nvidia-smi -i 0 --query-gpu=temperature.gpu --format=csv,noheader
*The ECC mode that the GPU is currently operating under
nvidia-smi -i 0 --query-gpu=ecc.mode.current --format=csv,noheader
*The power management status
nvidia-smi -i 0 --query-gpu=power.management --format=csv,noheader
*The last measured power draw for the entire board, in watts
nvidia-smi -i 0 --query-gpu=power.draw --format=csv,noheader
*The minimum and maximum value in watts that power limit can be set to.
nvidia-smi -i 0 --query-gpu=power.min_limit,power.max_limit --format=csv
* For management: ::
*Set persistence mode, When persistence mode is enabled the NVIDIA driver remains loaded even when no active clients, DISABLED by default
nvidia-smi -i 0 -pm 1
*Disabled ECC support for GPU. Toggle ECC support, A flag that indicates whether ECC support is enabled, need to use --query-gpu=ecc.mode.pending to check. Reboot required.
nvidia-smi -i 0 -e 0
*Reset the ECC volatile/aggregate error counters for the target GPUs
nvidia-smi -i 0 -p 0/1
*Set MODE for compute applications, query with --query-gpu=compute_mode
nvidia-smi -i 0 -c 0/1/2/3
*Trigger reset of the GPU.
nvidia-smi -i 0 -r
*Enable or disable Accounting Mode, statistics can be calculated for each compute process running on the GPU, query with -query-gpu=accounting.mode
nvidia-smi -i 0 -am 0/1
*Specifies maximum power management limit in watts, query with --query-gpu=power.limit.
nvidia-smi -i 0 -pl 200
**Installing CUDA example applications**
The cuda-samples-7-0 pkgs include some CUDA examples which can help uses to know how to use cuda.For a node which only cuda runtime libraries installed, the following command can be used to install cuda-samples package. ::
apt-get install cuda-samples-7-0 -y
After cuda-sample-7-0 has been installed, go to /usr/local/cuda-7.0/samples to build the examples. See this link https://developer.nvidia.com/ for more information. Or, you can simply run the make command under dir /usr/local/cuda-7.0/samples to build all the tools.
The following command can be used to build the deviceQuery tool in the cuda samples directory: ::
# pwd
/usr/local/cuda-7.0/samples
# make -C 1_Utilities/deviceQuery
make: Entering directory `/usr/local/cuda-7.0/samples/1_Utilities/deviceQuery'
/usr/local/cuda-7.0/bin/nvcc -ccbin g++ -I../../common/inc -m64 -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o deviceQuery.o -c deviceQuery.cpp
/usr/local/cuda-7.0/bin/nvcc -ccbin g++ -m64 -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o deviceQuery deviceQuery.o
mkdir -p ../../bin/ppc64le/linux/release
cp deviceQuery ../../bin/ppc64le/linux/release
make: Leaving directory `/usr/local/cuda-7.0/samples/1_Utilities/deviceQuery'
The verification results from this example on a test node were: ::
# pwd
/usr/local/cuda-7.0/samples
# bin/ppc64le/linux/release/deviceQuery
bin/ppc64le/linux/release/deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 4 CUDA Capable device(s)
Device 0: "Tesla K80"
CUDA Driver Version / Runtime Version 7.0 / 7.0
CUDA Capability Major/Minor version number: 3.7
Total amount of global memory: 11520 MBytes (12079136768 bytes)
(13) Multiprocessors, (192) CUDA Cores/MP: 2496 CUDA Cores
GPU Max Clock rate: 824 MHz (0.82 GHz)
Memory Clock rate: 2505 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 1572864 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Enabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 3 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >
Device 1: "Tesla K80"
CUDA Driver Version / Runtime Version 7.0 / 7.0
......

3
docs/source/advanced/gpu/index.rst
View File

@ -4,5 +4,4 @@ GPUs
.. toctree::
:maxdepth: 2
cuda_rhel.rst
cuda_ubuntu.rst
nvidia/index.rst

22
docs/source/advanced/gpu/nvidia/deploy_cuda_node.rst Normal file
View File

@ -0,0 +1,22 @@
Deploy CUDA nodes
=================
Diskful
-------
* To provision diskful nodes using osimage ``rhels7.2-ppc64le-install-cudafull``: ::
nodeset <noderange> osimage=rhels7.2-ppc64le-install-cudafull
rsetboot <noderange> net
rpower <noderange> boot
Diskless
--------
* To provision diskless nodes using osimage ``rhels7.2-ppc64le-netboot-cudafull``: ::
nodeset <noderange> osimage=rhels7.2-ppc64le-netboot-cudafull
rsetboot <noderange> net
rpower <noderange> boot

19
docs/source/advanced/gpu/nvidia/index.rst Normal file
View File

@ -0,0 +1,19 @@
NVIDIA CUDA
===========
CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model created by NVIDIA. It can be used to increase computing performance by leveraging the Graphics Processing Units (GPUs).
For more information, see NVIDIAs website: https://developer.nvidia.com/cuda-zone
xCAT supports CUDA installation for Ubuntu 14.04.3 and RHEL 7.2LE on PowerNV (Non-Virtualized) for both diskful and diskless nodes.
Within the NVIDIA CUDA Toolkit, installing the ``cuda`` package will install both the ``cuda-runtime`` and the ``cuda-toolkit``. The ``cuda-toolkit`` is intended for developing CUDA programs and monitoring CUDA jobs. If your particular installation requires only running GPU jobs, it's recommended to install only the ``cuda-runtime`` package.
.. toctree::
:maxdepth: 2
repo/index.rst
osimage/index.rst
deploy_cuda_node.rst
verify_cuda_install.rst
management.rst

107
docs/source/advanced/gpu/nvidia/management.rst Normal file
View File

@ -0,0 +1,107 @@
GPU Management and Monitoring
=============================
The ``nvidia-smi`` command provided by NVIDIA can be used to manage and monitor GPUs enabled Compute Nodes. In conjunction with the xCAT``xdsh`` command, you can easily manage and monitor the entire set of GPU enabled Compute Nodes remotely from the Management Node.
Example: ::
# xdsh <noderange> "nvidia-smi -i 0 --query-gpu=name,serial,uuid --format=csv,noheader"
node01: Tesla K80, 0322415075970, GPU-b4f79b83-c282-4409-a0e8-0da3e06a13c3
...
**Note: The following commands are provided as convenience.** *Always consult the nvidia-smi manpage for the latest supported functions.*
Management
----------
Some useful ``nvidia-smi`` example commands for management.
* Set persistence mode, When persistence mode is enabled the NVIDIA driver remains loaded even when no active clients, DISABLED by default::
nvidia-smi -i 0 -pm 1
* Disabled ECC support for GPU. Toggle ECC support, A flag that indicates whether ECC support is enabled, need to use --query-gpu=ecc.mode.pending to check [Reboot required]::
nvidia-smi -i 0 -e 0
* Reset the ECC volatile/aggregate error counters for the target GPUs::
nvidia-smi -i 0 -p 0/1
* Set MODE for compute applications, query with --query-gpu=compute_mode::
nvidia-smi -i 0 -c 0/1/2/3
* Trigger reset of the GPU ::
nvidia-smi -i 0 -r
* Enable or disable Accounting Mode, statistics can be calculated for each compute process running on the GPU, query with -query-gpu=accounting.mode::
nvidia-smi -i 0 -am 0/1
* Specifies maximum power management limit in watts, query with --query-gpu=power.limit ::
nvidia-smi -i 0 -pl 200
Monitoring
----------
Some useful ``nvidia-smi`` example commands for monitoring.
* The number of NVIDIA GPUs in the system ::
nvidia-smi --query-gpu=count --format=csv,noheader
* The version of the installed NVIDIA display driver ::
nvidia-smi -i 0 --query-gpu=driver_version --format=csv,noheader
* The BIOS of the GPU board ::
nvidia-smi -i 0 --query-gpu=vbios_version --format=csv,noheader
* Product name, serial number and UUID of the GPU::
nvidia-smi -i 0 --query-gpu=name,serial,uuid --format=csv,noheader
* Fan speed::
nvidia-smi -i 0 --query-gpu=fan.speed --format=csv,noheader
* The compute mode flag indicates whether individual or multiple compute applications may run on the GPU. (known as exclusivity modes) ::
nvidia-smi -i 0 --query-gpu=compute_mode --format=csv,noheader
* Percent of time over the past sample period during which one or more kernels was executing on the GPU::
nvidia-smi -i 0 --query-gpu=utilization.gpu --format=csv,noheader
* Total errors detected across entire chip. Sum of device_memory, register_file, l1_cache, l2_cache and texture_memory ::
nvidia-smi -i 0 --query-gpu=ecc.errors.corrected.aggregate.total --format=csv,noheader
* Core GPU temperature, in degrees C::
nvidia-smi -i 0 --query-gpu=temperature.gpu --format=csv,noheader
* The ECC mode that the GPU is currently operating under::
nvidia-smi -i 0 --query-gpu=ecc.mode.current --format=csv,noheader
* The power management status::
nvidia-smi -i 0 --query-gpu=power.management --format=csv,noheader
* The last measured power draw for the entire board, in watts::
nvidia-smi -i 0 --query-gpu=power.draw --format=csv,noheader
* The minimum and maximum value in watts that power limit can be set to ::
nvidia-smi -i 0 --query-gpu=power.min_limit,power.max_limit --format=csv

11
docs/source/advanced/gpu/nvidia/osimage/index.rst Normal file
View File

@ -0,0 +1,11 @@
Create osimage definitions
==========================
Generate ``osimage`` definitions to provison the compute nodes with the NVIDIA CUDA toolkit installed.
.. toctree::
:maxdepth: 2
rhels.rst
ubuntu.rst
postscripts.rst

47
docs/source/advanced/gpu/nvidia/osimage/postscripts.rst Normal file
View File

@ -0,0 +1,47 @@
Postscripts
===========
The following sections demonstrates how to use xCAT to configure post-installation steps
Setting PATH and LD_LIBRARY_PATH
--------------------------------
NVIDIA recommends various post-installation actions that should be performed to properly configure the nodes. A sample script is provided by xCAT for this purpose ``config_cuda`` and can be modified to fit your specific installation.
The ``config_cuda`` script sets the PATH and LD_LIBRARY_PATH: ::
#!/bin/sh
# set the paths required for cuda7.5
CUDA_VER="cuda-7.5"
FILENAME="/etc/profile.d/xcat-${CUDA_VER}.sh"
echo "export PATH=/usr/local/${CUDA_VER}/bin:\$PATH" > ${FILENAME}
echo "export LD_LIBRARY_PATH=/usr/local/${CUDA_VER}/lib64:\$LD_LIBRARY_PATH" >> ${FILENAME}
Add this script to your node object using the chdef command: ::
chdef -t node -o <noderange> -p postscripts=/install/postscripts/config_cuda
Setting GPU Configurations
--------------------------
NVIDIA allows for changing GPU attributes using the ``nvidia-smi`` commands. These settings do not persist when a compute node is rebooted. One way set these attributes is to use an xCAT postscript to set the values every time the node is rebooted.
* Set the power limit to 175W: ::
# set the power limit to 175W
nvidia-smi -pl 175
* Set the GPUs to persistence mode to increase performance: ::
# nvidia-smi -pm 1
Enabled persistence mode for GPU 0000:03:00.0.
Enabled persistence mode for GPU 0000:04:00.0.
Enabled persistence mode for GPU 0002:03:00.0.
Enabled persistence mode for GPU 0002:04:00.0.
All done.

115
docs/source/advanced/gpu/nvidia/osimage/rhels.rst Normal file
View File

@ -0,0 +1,115 @@
RHEL 7.2 LE
===========
Diskful images
---------------
The following examples will create diskful images for ``cudafull`` and ``cudaruntime``. The osimage definitions will be created from the base ``rhels7.2-ppc64le-install-compute`` osimage.
xCAT provides a sample package list files for CUDA. You can find them at:
* ``/opt/xcat/share/xcat/instal/rh/cudafull.rhels7.ppc64le.pkglist``
* ``/opt/xcat/share/xcat/instal/rh/cudaruntime.rhels7.ppc64le.pkglist``
**[diskful note]**: There is a requirement to reboot the machine after the CUDA drivers are installed. To satisfy this requirement, the CUDA software is installed in the ``pkglist`` attribute of the osimage definition where the reboot happens after the Operating System is installed.
cudafull
^^^^^^^^
#. Create a copy of the ``install-compute`` image and label it ``cudafull``: ::
lsdef -t osimage -z rhels7.2-ppc64le-install-compute \
| sed 's/install-compute:/install-cudafull:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``pkgdir`` attribute: ::
chdef -t osimage -o rhels7.2-ppc64le-install-cudafull -p pkgdir=/install/cuda-repo
#. Use the provided ``cudafull`` pkglist to install the CUDA packages: ::
chdef -t osimage -o rhels7.2-ppc64le-install-cudafull \
pkglist=/opt/xcat/share/xcat/instal/rh/cudafull.rhels7.ppc64le.pkglist
cudaruntime
^^^^^^^^^^^
#. Create a copy of the ``install-compute`` image and label it ``cudaruntime``: ::
lsdef -t osimage -z rhels7.2-ppc64le-install-compute \
| sed 's/install-compute:/install-cudaruntime:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``pkgdir`` attribute: ::
chdef -t osimage -o rhels7.2-ppc64le-install-cudaruntime -p pkgdir=/install/cuda-repo
#. Use the provided ``cudaruntime`` pkglist to install the CUDA packages: ::
chdef -t osimage -o rhels7.2-ppc64le-install-cudaruntime \
pkglist=/opt/xcat/share/xcat/instal/rh/cudaruntime.rhels7.ppc64le.pkglist
Diskless images
---------------
The following examples will create diskless images for ``cudafull`` and ``cudaruntime``. The osimage definitions will be created from the base ``rhels7.2-ppc64le-netboot-compute`` osimage.
xCAT provides a sample package list files for CUDA. You can find them at:
* ``/opt/xcat/share/xcat/netboot/rh/cudafull.rhels7.ppc64le.otherpkgs.pkglist``
* ``/opt/xcat/share/xcat/netboot/rh/cudaruntime.rhels7.ppc64le.otherpkgs.pkglist``
**[diskless note]**: For diskless images, the requirement for rebooting the machine is not applicable because the images is loaded on each reboot. The install of the CUDA packages is required to be done in the ``otherpkglist`` **NOT** the ``pkglist``.
cudafull
^^^^^^^^
#. Create a copy of the ``netboot-compute`` image and label it ``cudafull``: ::
lsdef -t osimage -z rhels7.2-ppc64le-netboot-compute \
| sed 's/netboot-compute:/netboot-cudafull:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``otherpkgdir`` attribute: ::
chdef -t osimage -o rhels7.2-ppc64le-netboot-cudafull otherpkgdir=/install/cuda-repo
#. Add the provided ``cudafull`` otherpkglist.pkglist file to install the CUDA packages: ::
chdef -t osimage -o rhels7.2-ppc64le-netboot-cudafull \
otherpkglist=/opt/xcat/share/xcat/netboot/rh/cudafull.rhels7.ppc64le.otherpkgs.pkglist
#. Generate the image: ::
genimage rhels7.2-ppc64le-netboot-cudafull
#. Package the image: ::
packimage rhels7.2-ppc64le-netboot-cudafull
cudaruntime
^^^^^^^^^^^
#. Create a copy of the ``netboot-compute`` image and label it ``cudaruntime``: ::
lsdef -t osimage -z rhels7.2-ppc64le-netboot-compute \
| sed 's/netboot-compute:/netboot-cudaruntime:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``otherpkgdir`` attribute: ::
chdef -t osimage -o rhels7.2-ppc64le-netboot-cudaruntime otherpkgdir=/install/cuda-repo
#. Add the provided ``cudaruntime`` otherpkglist.pkglist file to install the CUDA packages: ::
chdef -t osimage -o rhels7.2-ppc64le-netboot-cudaruntime \
otherpkglist=/opt/xcat/share/xcat/netboot/rh/cudaruntime.rhels7.ppc64le.otherpkgs.pkglist
#. Generate the image: ::
genimage rhels7.2-ppc64le-netboot-cudaruntime
#. Package the image: ::
packimage rhels7.2-ppc64le-netboot-cudaruntime

158
docs/source/advanced/gpu/nvidia/osimage/ubuntu.rst Normal file
View File

@ -0,0 +1,158 @@
Ubuntu 14.04.3
==============
Diskful images
---------------
The following examples will create diskful images for ``cudafull`` and ``cudaruntime``. The osimage definitions will be created from the base ``ubuntu14.04.3-ppc64el-install-compute`` osimage.
xCAT provides a sample package list files for CUDA. You can find them at:
* ``/opt/xcat/share/xcat/install/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist``
* ``/opt/xcat/share/xcat/install/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist``
**[diskful note]**: There is a requirement to reboot the machine after the CUDA drivers are installed. To satisfy this requirement, the CUDA software is installed in the ``pkglist`` attribute of the osimage definition where the reboot happens after the Operating System is installed.
cudafull
^^^^^^^^
#. Create a copy of the ``install-compute`` image and label it ``cudafull``: ::
lsdef -t osimage -z ubuntu14.04.3-ppc64el-install-compute \
| sed 's/install-compute:/install-cudafull:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``pkgdir`` attribute.
If your Management Node IP is 10.0.0.1, the URL for the repo would be ``http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local``, add it to the pkgdir::
chdef -t osimage -o ubuntu14.04.3-ppc64el-install-cudafull \
-p pkgdir=http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local
**TODO:** Need to add Ubuntu Port? "http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main"
#. Use the provided ``cudafull`` pkglist to install the CUDA packages: ::
chdef -t osimage -o ubuntu14.04.3-ppc64el-install-cudafull \
pkglist=/opt/xcat/share/xcat/install/ubuntu/cudafull.ubuntu14.04.3.ppc64el.pkglist
cudaruntime
^^^^^^^^^^^
#. Create a copy of the ``install-compute`` image and label it ``cudaruntime``: ::
lsdef -t osimage -z ubuntu14.04.3-ppc64el-install-compute \
| sed 's/install-compute:/install-cudaruntime:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``pkgdir`` attribute:
If your Management Node IP is 10.0.0.1, the URL for the repo would be ``http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local``, add it to the pkgdir::
chdef -t osimage -o ubuntu14.04.3-ppc64el-install-cudaruntime \
-p pkgdir=http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local
**TODO:** Need to add Ubuntu Port? "http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main"
#. Use the provided ``cudaruntime`` pkglist to install the CUDA packages: ::
chdef -t osimage -o ubuntu14.04.3-ppc64el-install-cudaruntime \
pkglist=/opt/xcat/share/xcat/install/ubuntu/cudaruntime.ubuntu14.04.3.ppc64el.pkglist
Diskless images
---------------
The following examples will create diskless images for ``cudafull`` and ``cudaruntime``. The osimage definitions will be created from the base ``ubuntu14.04.3-ppc64el-netboot-compute`` osimage.
xCAT provides a sample package list files for CUDA. You can find them at:
* ``/opt/xcat/share/xcat/netboot/rh/cudafull.rhels7.ppc64le.otherpkgs.pkglist``
* ``/opt/xcat/share/xcat/netboot/rh/cudaruntime.rhels7.ppc64le.otherpkgs.pkglist``
**[diskless note]**: For diskless images, the requirement for rebooting the machine is not applicable because the images is loaded on each reboot. The install of the CUDA packages is required to be done in the ``otherpkglist`` **NOT** the ``pkglist``.
cudafull
^^^^^^^^
#. Create a copy of the ``netboot-compute`` image and label it ``cudafull``: ::
lsdef -t osimage -z ubuntu14.04.3-ppc64el-netboot-compute \
| sed 's/netboot-compute:/netboot-cudafull:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``otherpkgdir`` attribute.
If your Management Node IP is 10.0.0.1, the URL for the repo would be ``http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local``, add it to the ``otherpkgdir``::
chdef -t osimage -o ubuntu14.04.3-ppc64el-netboot-cudafull \
otherpkgdir=http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local
#. Add the provided ``cudafull`` otherpkg.pkglist file to install the CUDA packages: ::
chdef -t osimage -o ubuntu14.04.3-ppc64el-netboot-cudafull \
otherpkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudafull.otherpkgs.pkglist
**TODO:** Need to add Ubuntu Port? "http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main"
#. Verify that ``acpid`` is installed on the Management Node or on the Ubuntu host where you are generating the diskless image: ::
apt-get install -y acpid
#. Generate the image: ::
genimage ubuntu14.04.3-ppc64el-netboot-cudafull
#. Package the image: ::
packimage ubuntu14.04.3-ppc64el-netboot-cudafull
cudaruntime
^^^^^^^^^^^
#. Create a copy of the ``netboot-compute`` image and label it ``cudaruntime``: ::
lsdef -t osimage -z ubuntu14.04.3-ppc64el-netboot-compute \
| sed 's/netboot-compute:/netboot-cudaruntime:/' \
| mkdef -z
#. Add the CUDA repo created in the previous step to the ``otherpkgdir`` attribute.
If your Management Node IP is 10.0.0.1, the URL for the repo would be ``http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local``, add it to the ``otherpkgdir``::
chdef -t osimage -o ubuntu14.04.3-ppc64el-netboot-cudaruntime \
otherpkgdir=http://10.0.0.1/install/cuda-repo/ppc64/var/cuda-repo-7-5-local
#. Add the provided ``cudaruntime`` otherpkg.pkglist file to install the CUDA packages: ::
chdef -t osimage -o ubuntu14.04.3-ppc64el-netboot-cudaruntime \
otherpkglist=/opt/xcat/share/xcat/netboot/ubuntu/cudaruntime.otherpkgs.pkglist
**TODO:** Need to add Ubuntu Port? "http://ports.ubuntu.com/ubuntu-ports trusty main,http://ports.ubuntu.com/ubuntu-ports trusty-updates main"
#. Verify that ``acpid`` is installed on the Management Node or on the Ubuntu host where you are generating the diskless image: ::
apt-get install -y acpid
#. Generate the image: ::
genimage ubuntu14.04.3-ppc64el-netboot-cudaruntime
#. Package the image: ::
packimage ubuntu14.04.3-ppc64el-netboot-cudaruntime
Install NVIDIA Management Library (optional)
--------------------------------------------
See https://developer.nvidia.com/nvidia-management-library-nvml for more information.
The .run file can be downloaded from NVIDIAs website and placed into the ``/install/postscripts`` directory on the Management Node.
To enable installation of the management library after the node is install, add the runfile to the ``postbootscripts`` attribute for the nodes: ::
chmod +x /install/postscripts/<gpu_deployment_kit>.run
chdef -t node -o <noderange> -p postbootscripts=<gpu_deployment_kit>.run \
--silent --installdir=<your_desired_install_dir>

12
docs/source/advanced/gpu/nvidia/repo/index.rst Normal file
View File

@ -0,0 +1,12 @@
Create CUDA software repository
===============================
The NVIDIA CUDA Toolkit is available to download at http://developer.nvidia.com/cuda-downloads.
Download the toolkit and prepare the software repository on the xCAT Management Node to server the NVIDIA CUDA files.
.. toctree::
:maxdepth: 2
rhels.rst
ubuntu.rst

31
docs/source/advanced/gpu/nvidia/repo/rhels.rst Normal file
View File

@ -0,0 +1,31 @@
RHEL 7.2 LE
===========
#. Create a repository on the MN node installing the CUDA Toolkit: ::
# For cuda toolkit name: /root/cuda-repo-rhel7-7-5-local-7.5-18.ppc64le.rpm
# extract the contents from the rpm
mkdir -p /tmp/cuda
cd /tmp/cuda
rpm2cpio /root/cuda-repo-rhel7-7-5-local-7.5-18.ppc64le.rpm | cpio -i -d
# Create the repo directory under xCAT /install dir
mkdir -p /install/cuda-repo/ppc64le
cp -r /tmp/cuda/var/cuda-repo-7-5-local /install/cuda-repo/ppc64le/
# Create the yum repo files
createrepo /install/cuda-repo/ppc64le
#. The NVIDIA CUDA Toolkit contains rpms that have dependencies on other external packages (such as ``DKMS``). These are provided by EPEL. It's up to the system administrator to obtain the dependency packages and add those to the ``cuda-deps`` directory: ::
mkdir -p /install/cuda-repo/cuda-deps
cd /install/cuda-repo/cuda-deps
# Copy the DKMS rpm to this directory
ls -ltr /install/cuda-repo/cuda-deps
-rw-r--r-- 1 root root 79048 Oct 5 10:58 dkms-2.2.0.3-30.git.7c3e7c5.el7.noarch.rpm
# Execute createrepo in this directory
createrepo .

36
docs/source/advanced/gpu/nvidia/repo/ubuntu.rst Normal file
View File

@ -0,0 +1,36 @@
Ubuntu 14.04.3
==============
NVIDIA supports two types of debian repositories that can be used to install Cuda Toolkit: **local** and **network**. You can download the installers from https://developer.nvidia.com/cuda-downloads.
local
-----
A local package repo will contain all of the cuda packages.
Extract the cuda packages into ``/install/cuda-repo/ppc64le``: ::
# For cuda toolkit name: /root/cuda-repo-ubuntu1404-7-5-local_7.5-18_ppc64el.deb
# Create the repo directory under xCAT /install dir
mkdir -p /install/cuda-repo/ppc64le
dpkg -x /root/cuda-repo-ubuntu1404-7-5-local_7.5-18_ppc64el.deb /install/cuda-repo/ppc64le
network
-------
The online package repo provides a source list entry pointing to a URL containing the CUDA packages. This can be used directly on the Compute Nodes.
The ``sources.list`` entry may look similar to: ::
deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1410/ppc64el /
Authorize the CUDA repo
-----------------------
In order to access the CUDA repository you must import the CUDA GPGKEY into the ``apt_key`` trust list. xCAT provides a sample postscript ``/install/postscripts/addcudakey`` to help with this task: ::
chdef -t node -o <noderange> -p postscripts=addcudakey

80
docs/source/advanced/gpu/nvidia/verify_cuda_install.rst Normal file
View File

@ -0,0 +1,80 @@
Verify CUDA Installation
========================
**The following verification steps only apply to the ``cudafull`` installations.**
#. Verify driver version by looking at: ``/proc/driver/nvidia/version``: ::
# cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX ppc64le Kernel Module 352.39 Fri Aug 14 17:10:41 PDT 2015
GCC version: gcc version 4.8.5 20150623 (Red Hat 4.8.5-4) (GCC)
#. Verify the CUDA Toolkit version ::
# nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Tue_Aug_11_14:31:50_CDT_2015
Cuda compilation tools, release 7.5, V7.5.17
#. Verify running CUDA GPU jobs by compiling the samples and executing the ``deviceQuery`` or ``bandwidthTest`` programs.
* Compile the samples:
**[RHEL]:** ::
cd ~/
cuda-install-samples-7.5.sh .
cd NVIDIA_CUDA-7.5_Samples
make
**[Ubuntu]:** ::
cd ~/
apt-get install cuda-samples-7-0 -y
cd /usr/local/cuda-7.0/samples
make
* Run the ``deviceQuery`` sample: ::
# ./bin/ppc64le/linux/release/deviceQuery
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 4 CUDA Capable device(s)
Device 0: "Tesla K80"
CUDA Driver Version / Runtime Version 7.5 / 7.5
CUDA Capability Major/Minor version number: 3.7
Total amount of global memory: 11520 MBytes (12079136768 bytes)
(13) Multiprocessors, (192) CUDA Cores/MP: 2496 CUDA Cores
GPU Max Clock rate: 824 MHz (0.82 GHz)
Memory Clock rate: 2505 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 1572864 bytes
............
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 7.5, CUDA Runtime Version = 7.5, NumDevs = 4, Device0 = Tesla K80, Device1 = Tesla K80, Device2 = Tesla K80, Device3 = Tesla K80
Result = PASS
* Run the ``bandwidthTest`` sample: ::
# ./bin/ppc64le/linux/release/bandwidthTest
[CUDA Bandwidth Test] - Starting...
Running on...
Device 0: Tesla K80
Quick Mode
Host to Device Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 7765.1
Device to Host Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 7759.6
Device to Device Bandwidth, 1 Device(s)
PINNED Memory Transfers
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 141485.3
Result = PASS
NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.