docs(rfc): add sandbox resource requirements proposal#1360
Conversation
elezar
requested review from
a team,
derekwaynecarr,
maxamillion and
mrunalp
as code owners
Signed-off-by: Evan Lezar <elezar@nvidia.com>
elezar
force-pushed
the
1338-sandbox-resource-requirements/elezar
branch
from
96945aa to
1aafb69
Compare
| The CLI should not expose a JSON flag for `resource_requirements`. Common | ||
| portable requests should use typed flags such as CPU, memory, and GPU-count | ||
| flags, and SDK/API callers should use the typed protobuf messages directly. | ||
| JSON-formatted CLI input is reserved for driver-specific configuration. |
There was a problem hiding this comment.
This seems right to me. NemoClaw team needs basic mem/cpu requests, so I started to implement this here, #1376.
| repeated GenericResourceRequirement extensions = 100; | ||
| } | ||
|
|
||
| message ComputeResourceRequirements { |
There was a problem hiding this comment.
I’m good with this.
We may want ephemeral storage in future, but am happy to defer that now.
I am not aware of any use case right now that would demand hugepages.
Right now pid limiting in k8s is a node level pod setting that is per pod cgroup enforced, if we wanted to expose a pids.max. Ultimately we need our pid limiting to be cgroup enforced but I think having that as settable can come later.
|
|
||
| | Driver | Realization | | ||
| |---|---| | ||
| | Kubernetes | Populate pod container `resources.requests.cpu`, `resources.limits.cpu`, `resources.requests.memory`, and `resources.limits.memory`. | |
There was a problem hiding this comment.
if and when we separate the proxy into a separate pod, we may need to revisit this slightly to either have a fixed proxy overhead or enable it to have a separate resource configuration. We can explore that when we get there as this rfc improves the current model.
derekwaynecarr
left a comment
derekwaynecarr
left a comment
There was a problem hiding this comment.
This is a good next step. I wanted to let @mrunalp take a look as well, but this is LGTM for me.
| ```shell | ||
| openshell sandbox create \ | ||
| --driver-config-json '{"kubernetes.openshell.ai":{"nodeSelector":{"accelerator":"nvidia"}}}' | ||
| ``` |
There was a problem hiding this comment.
This will be really useful across all our drivers.
3 participants
Summary
Add RFC 0004 proposing a typed sandbox resource requirements model for CPU, memory, GPUs, and future resource domains. The RFC separates portable resource requirements from driver/platform-specific configuration and realization, and includes concrete realization examples for Kubernetes, Docker, Podman, and VM drivers.
Related Issue
Related to #1338 and #1340.
Changes
rfc/0004-sandbox-resource-requirements/README.md.SandboxSpec.resource_requirementswith compute, device, dataset, and extension domains.--driver-config-json, mapped toSandboxTemplate.driver_config.SandboxTemplate.resourcespassthrough.Testing
mise run pre-commitpassesChecklist