Federated Cloud GPGPU
GPGPU on Federated Cloud (PRE-PRODUCTION!)
GPGPU is available on selected sites of the EGI Federated Cloud, described in the table below:
|Site Name||Supported VOs||OCCI endpoint||OpenStack endpoint||GPU templates/flavors|
|IISAS-GPU||fedcloud.egi.eu, ops, dteam, moldyngrid, enmr.eu, vo.lifewatch.eu||https://nova3.ui.savba.sk:8787/occi1.1/||https://keystone3.ui.savba.sk:5000/v2.0||gpu1cpu6 (1GPU + 6 CPU cores), gpu2cpu12 (2GPU +12 CPU cores)|
Instantiate GPGPUs VMs
Creating VMs with GPGPUs is done as with every other VM, you just need to select the appropriate template. First of all, set up your interface following the CLI setup guide.
Then you can discover which templates do support GPGPU by describing them:
occi --endpoint $OCCI_ENDPOINT \ --auth x509 --user-cred $X509_USER_PROXY --voms \ --action describe --resource resource_tpl
This will show you the list of templates with a short description, look for the ones with
gpu in the description listed in the table above, e.g.:
[[ http://schemas.openstack.org/template/resource#f0cd78ab-10a0-4350-a6cb-5f3fdd6e6294 ]] title: Flavor: gpu1cpu6 term: f0cd78ab-10a0-4350-a6cb-5f3fdd6e6294 location: /mixin/f0cd78ab-10a0-4350-a6cb-5f3fdd6e6294/
You can start any VM, for testing you may use one with CUDA support like the EGI-Cuda appliance. Start your VM with the selected templates, make sure that you add proper context information to be able to login (check the fedcloud FAQ for more information):
RES_TPL=<set this to the selected resource template> OS_TPL=<set this to the selected VM image> occi --endpoint $OCCI_ENDPOINT \ --auth x509 --user-cred $X509_USER_PROXY --voms \ --action create --resource compute \ --mixin $OS_TPL --mixin $RES_TPL \ --attribute occi.core.title="Testing GPU" \ --context <add here any contextualization>
If the available VA do not suite your needs, you can install NVIDIA driver and CUDA toolkit to a VM. They are available at http://www.nvidia.com/Download/index.aspx and https://developer.nvidia.com/cuda-downloads. See NVIDIA_CUDA_installer for a sample script to install on debian based VMs
Create your own GPGPU Virtual Appliances
You can create and upload your Virtual Appliances with your applications that can be replicated to sites once endorsed by a VO.
We recommend using a tool like packer for creating the images. Check CUDA packer file of the VMI endorsement repo for a working configuration to create such Virtual Appliance. Starting from the example, you can add new
provisioners to install your applications.