sig-node e2e tests machine hardware requirements

[ffromani]

sig-node owns a set of features related to exposing and using hardware details which require some hardware features to exercise the code. Examples are exclusive CPU allocation (cpumanager), device allocation (device manager), NUMA alignment (topology manager), NUMA alignment considering distances between NUMA zones (topology manager).

Note: some requirement overlap. Easy example: a powerful high end (at time of writing) server CPU can have at the same time multi core count, exposing multiple NUMA nodes, and have split L3, satisfying in one go all cpumanager requirements

Hardware requirements, driven by feature, rationale

• cpumanager (GA): x86_64, arm: machine with at least 4 cores exposed, better like 16 or more. We need more cores to be able to run different set of tests in different scenarios. 4 cores is the minimum to run basic test about the feature. We already got machines with 4 cores
• topology manager (GA): x86_64, arm: machine with 2 or more NUMA nodes. We need to align on NUMA nodes, so we need 2 or more nodes to begin with
• topology manager (KEP https://github.com/kubernetes/enhancements/issues/3545): x86_64, arm: machine with 4 or more NUMA nodes. We need to consider NUMA distances in allocation, so we need 4 or more nodes with different distances between each other
• topology manager (KEP https://github.com/kubernetes/enhancements/issues/4622): arm only?: machine with 9 or more NUMA nodes (!) like grace gpus
• cpumanager (KEP https://github.com/kubernetes/enhancements/pull/4810#): x86_64, arm: a machine with split (non-uniform) L3 cache, like epyc cpus
• cpumanager (KEP https://github.com/kubernetes/enhancements/issues/3327): x86_64, arm: a machine with multiple CPU sockets, in order to exercise alignment by socket.
• devicemanager (GA): x86_64, arm: hardware devices controlled by device plugins. Most common usecases are SRIOV cards and GPUs. One device is the bare minimum, we would like 2 or more to be able to use them also in the topology manager tests as deciding factor (see topology manager in this list)
• memorymanager (Beta, GA graduating): x86_64, arm: totally overlaps with topology manager requirements, listed here for the sake of completeness.

this list will be updated after more review of the ongoing sig-node features

[ffromani]

tagging some relevant sig-node people: @kannon92 @PiotrProkop @klueska

[ffromani]

slack thread for context: https://kubernetes.slack.com/archives/CCK68P2Q2/p1727202732284529

[ameukam]

cc @dims @upodroid @BenTheElder

[BenTheElder]

We have EC2 and GCE pretty well setup in particular at the moment, do any of the machine types available there meet your requirements?

Please make sure any new resources you use on any platform are handled by the kubernetes-sigs/boskos cleanup scripts. If you're using GCP projects / AWS accounts with VMs that should already work.

[ffromani]

@catblade kindly pointed out equinix donated cloud credits and their offering seems also interesting and maybe we can use it. Some CNCF TAGs already make use if it. This is the reference I got: https://github.com/cncf-tags/green-reviews-tooling/

[ameukam]

@ffromani why can't we use AWS EC2 instances to run those tests ?

[ameukam]

CPU architectures available:

GCP: https://cloud.google.com/compute/docs/cpu-platforms AWS:

• https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/cpu-options-supported-instances-values.html
• https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-types.html#instance-hardware-specs

[ffromani]

@ffromani why can't we use AWS EC2 instances to run those tests ?

I think we totally can, I'm not aware of any blocker. The efforts in this area have been somehow sparse, we're taking the chance of sig-node 1.32 planning to re-evaluate and improve the current state. Will review the GCP/AWS offerings and comment.

[BenTheElder]

@catblade kindly pointed out equinix donated cloud credits and their offering seems also interesting and maybe we can use it. Some CNCF TAGs already make use if it. This is the reference I got: https://github.com/cncf-tags/green-reviews-tooling/

Yes, however we generally are running critical infra on Kubernetes specific resource allocations, and we don't currently have a lot setup to manage this. For equinix SIG K8s Infra doesn't currently have observability into the amount of resources available and the spending trends which has bitten us in the past (see reports like https://kubernetes.slack.com/archives/CCK68P2Q2/p1727127173398879 for some of the others).

(@dims does have cs.k8s.io running on equinix currently, we also have some presence in DO and Azure but not as mature yet, and Fastly for CDN)

It would be easier if we can use one of vendors for which we already have tooling (like https://github.com/kubernetes-sigs/boskos) setup to avoid resource leaks etc.

Otherwise we need help to invest in and onboard new resource types, observability into utilization and remaining credits, etc

[ffromani]

review of the GCP offering: AMD: https://cloud.google.com/compute/docs/cpu-platforms#amd_processors We do have some epycs which seems to fit all the requirements. We need perhaps to doublecheck the > 2 NUMA nodes requirement which brings us to C2D or Tau-T2D instances which are surely good:

• c2d >= c2d-standard-32 <- note SMT is enabled here
• tau-t2d >= t2d-standard-32 <- note SMT is disabled here

considering the SMT support status c2d seems better for our purposes

Intel: https://cloud.google.com/compute/docs/cpu-platforms#intel_processors IIRC we already have lanes running on n1 or n2 instances, which should be good for everything but split L3 instances and > 2 NUMA instances:

• n2 >= n2-standard-32

AI: review current usage of intel n2 instances and comment here

[ameukam]

Can we also get the same analysis on AWS ? and possibly on Azure ?

[ffromani]

Can we also get the same analysis on AWS ? and possibly on Azure ?

yes, ongoing

[ffromani]

review of the AWS offering: Looking at the available instance types the best candidates (all factors considered, including budget-aware selection) seems to be "compute intensive" or "memory intensive".

compute intensive: Intel: c5 >= c5.12xlarge seems OK but is unlikely (docs seems more opaque) to be multi-numa and multi-socket. Also unclear if it is split-L3. AMD: managed to find some deep-dive information and c5a >= c5a.8xlarge seems OK. Noteworthy that these instances seems to DO HAVE split-L3 but needs to be verified if these are multi NUMA. Unlilkely these are multi-socket.

memory intensive: AMD: we have interesting CPUs EPYC 7571 which by the chatter in the internet seems to be AWS-specific versions of AMD 7601 SKUs which in turn do have split L3 but still unclear the multi-numa and multi-socket status. r5a >= r5a.16xlarge seems to be OK with the usual caveats

Intel: likewise, with the usual caveat, r5 >= r5.16xlarge seems to be OK

[ffromani]

so I need to figure out how's the multi-NUMA and multi-socket situation on both AWS and GCP.

Chances are high that if we require "high enough" cores, we exceed the silicon limits and then the cloud provider is forced to give us multi-socket instances (a single physical socket can only have so many cores), but this is and likely will be us second-guessing them.

In addition, need to review the current usage for periodic non-presubmit jobs. IIRC we have n1 or n2 intances, will check and comment here and crosslink old conversations.

Coming up next: azure.