[Networking] High-Performance Pod-to-Pod Communication

[janetkuo]

SHOULD: Utilize DRA for network resources to attach Pods to multiple network interfaces. This is the recommended way to expose specialized high-performance network interfaces to Pods, and their characteristics should be discoverable to enable informed scheduling or workload configuration.

[elieserr]

would rephrase the first part:

Pods should be able to discover and attach network devices trough DRA.

also maybe at the end

This is related to Hardware Topology Awareness section above.

[kannon92]

Utilize DRA for network resources to attach Pods to multiple network interfaces.

AFAIK DRA would require a driver so I'm curious what drivers actually do this.

I know of https://dranet.dev/docs/quick-start/ (cc @aojea).

Are there other solutions that are possible?

[aojea]

The Kubernetes networking community make a lot of discussion around this topic, and we decided to use DRA for anything multi network so we can standardize the ecosystem using common APIs, see KEP-4817: Resource Claim Status With Possible Standardized Network Interface Data
that is beta enabled by default in 1.34 and is targeting GA in 1.35.

To simplify even more the stack and help drivers implementers we also added the capability to move networks interfaces to the OCI runtime spec opencontainers/runtime-spec#1271 , so the ecosystem will allow to declarative attach the network interfaces, see runc and crun and NRI and CDI

Dranet is one example, there is also https://github.com/kubernetes-sigs/cni-dra-driver and I know other vendors are planning to use NRI or CDI once runc 1.4 is released with this new capability of attaching network interface, specially because it solves the problem of require privileged drivers. Even the multinetwork working group is using this https://github.com/kubernetes-sigs/multi-network.

The use of DRA provides truly vendor compatibility and portability based on standard and common APIs, my paper about Kubernetes Network Drivers goes more into detail in all this topics.

The point is that you do not need a specific driver to do this, ANY driver can do it, a GPU driver can also do the network attachment per example, so I do no think this should encourage the use of one or other particular project, just that implementations should use the standard APIs that exist (in this case in the ResourceClaim) so we allow application portability and standardization, that is the goal of the conformance programs.