Difference between revisions of "GridPP DIRAC For SiteAdmins"
(→GPU) |
|||
| Line 13: | Line 13: | ||
==GPU== | ==GPU== | ||
| + | |||
| + | GridPP DIRAC assumes that your GPUs have a dedicated queue per GPU type. <br> | ||
| + | A discussed in the technical meeting, for each GPU queue create an extension tag to the CE: | ||
| + | GPU_GridPP_<queue> with value <Software>_<Memory in GiB>[_EX] (note that queue may itself contain underscores). | ||
| + | |||
| + | Software: CUDA, OpenCL <br> | ||
| + | Memory: Integer rounded down to the nearest GiB, zero is OK <br> | ||
| + | _EX: runs in exclusive compute mode <br> | ||
| + | An example would be, here for a Nvidia Tesla K40 <br> | ||
| + | GPU_GridPP_my_amazing_queue CUDA_11 | ||
Revision as of 13:08, 21 November 2017
Pilot Role
Please remember to enable the pilot role for all VOs supported by the GridPP DIRAC instance.
If you (as a site admin) notice a distinct lack of jobs submitted via the GridPP DIRAC instance, please contact Simon and Daniela at Imperial.
VAC/Vcycle
To enable automatic configuration of your VAC/Vcycle site, please create the following extension for the service in GOCDB:
PILOT_DN_GridPP <user_data_proxy DN> PILOT_SE_GridPP <GridPP DIRAC SE NAME>
The update agent runs every 6 h. Please take this into account when doing updates.
Here is an example for the Imperial vcycle site.
GPU
GridPP DIRAC assumes that your GPUs have a dedicated queue per GPU type.
A discussed in the technical meeting, for each GPU queue create an extension tag to the CE:
GPU_GridPP_<queue> with value <Software>_<Memory in GiB>[_EX] (note that queue may itself contain underscores).
Software: CUDA, OpenCL
Memory: Integer rounded down to the nearest GiB, zero is OK
_EX: runs in exclusive compute mode
An example would be, here for a Nvidia Tesla K40
GPU_GridPP_my_amazing_queue CUDA_11
