Using the IMP for containment in a user instance #119
SteVwonder
started this conversation in
Design
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
In some cases, it would be beneficial for user instances to be able to use a subset of the IMP's functionality. In particular, a user instance may want to contain the resources of its individual jobs, but that containment mechanism requires root.
Two potential use cases:
For the network example: with Cray Aries, the NIC on the node has to be configured for each parallel job that is launched. In this case, the NIC resources allocated to the parallel job should be proportional to the size of the parallel job. It also does not appear possible to pre-allocate a bunch of NIC resources for the nested Flux instance and allow it to launch its jobs within those limits (although that is what I would push for for the Flux architecture).
For the power example: each job may be allocated a cumulative power limit. Nodes within that job may each start with the same power limit, and as the application runs, runtimes may want to shift power between nodes. This will require coordination between the runtime, Flux, and the IMP.
Beta Was this translation helpful? Give feedback.
All reactions