Multi-cluster policy
From Grid5000
An uniform way to manage the clusters of a site.
Contents |
Why a convention on the multi-cluster management?
Rationalize the configuration of the sites so:
- Users can handle the platform more efficiently
- Administrators can share their way to manage the platform
Convention policy
What is a cluster?
A cluster is a relatively homogeneous set of nodes:
- homogeneous in term of deployment (configuration and performance)
- homogeneous for user's experiments
This smooth definition allows to adapt to each case when new hardware is purchased.
Note: a not mandatory but good way to know when a new cluster appears is when a new value for the (site,nodemodel,cputype,cpufreq) couple is defined
User frontends
One user frontend must at least be available on each site.
Default requirements
All the user frontends must allow users to:
- authenticate themselves
- manage their data
- submit jobs
- access all the resources of the grid
Main user frontend
A main user frontend is elected on each site and must fulfilled additional requirements:
- be the main login frontend machine
- be the main job submission machine
- be the machine holding Kadeploy's binaries
The main user frontend helps improving platform's understanding and scripting comfort.
Services
| Todo |
|---|---|
Load balancing may be part of this draft one day, but its is not required for the OAR2 migration | |
Job submission
Users may be able to reserve from the user frontends:
- any resources of the local site with OAR, to improve job allocation
- any resources of the grid with OAR Grid
Deployment
User may be able to deploy any of the resources of the local site from the main user frontend.
Procedures
Receiving new compute nodes
- During a meeting of the CT, the site presents its new hardware and its will of integration
- Then the members of the CT study if the demand can be fulfilled without harming platform's uniformity
- At the end of the meeting, consensus is reached about integrating the new hardware as a new cluster or as an extension
Retiring old compute nodes
|
| Todo |
|---|---|
To draft one day, but not required for the OAR2 migration | |
