This PR makes the NodePrepareResources() and NodeUnprepareResource()
calls of the kubeletplugin API for DynamicResourceAllocation
symmetrical. It wasn't clear how one would use the set of CDIDevices
passed back in the NodeUnprepareResource() of the v1alpha1 API, and the
new API now passes back the full ResourceHandle that was originally
passed to the Prepare() call. Passing the ResourceHandle is strictly
more informative and a plugin could always (re)derive the set of
CDIDevice from it.
This is a breaking change, but this release is scheduled to break
multiple APIs for DynamicResourceAllocation, so it makes sense to do
this now instead of later.
Signed-off-by: Kevin Klues <kklues@nvidia.com>
* Update pod_container_manager_linux.go
This is a simple optimization to reduce repeated invoking of the GetPodContainerName function.
* Update pod_container_manager_linux.go
将podContainerName, _ := m.GetPodContainerName(pod)更靠近使用podcontainerName变量的位置
Right now, the v1alpha1 API only passes enough information for one plugin to
process a claim, but the v1alpha2 API will allow for multiple plugins to
process a claim. This commit prepares the code for this upcoming change.
Signed-off-by: Kevin Klues <kklues@nvidia.com>
* add timeouts for communication with dra plugin
* move timeout constant to k8s.io/kubernetes/pkg/kubelet/cm/util
* move settings of timeout to pkg/kubelet/plugin/dra/plugin/client.go
* remove timeout constant
The checkpointing mechanism will repopulate DRA Manager in-memory cache on kubelet restart.
This will ensure that the information needed by the PodResources API is available across
a kubelet restart.
The ClaimInfoState struct represent the DRA Manager in-memory cache state in checkpoint.
It is embedd in the ClaimInfo which also include the annotation field. The separation between
the in-memory cache and the cache state in the checkpoint is so we won't be tied to the in-memory
cache struct which may change in the future. In the ClaimInfoState we save the minimal required fields
to restore the in-memory cache.
Signed-off-by: Moshe Levi <moshele@nvidia.com>
With Topology Manager enabled by default, we no longer need
`resourceAllocator` as Topology Manager serves as the main
PodAdmitHandler completely responsible for admission check
based on hints received from the hintProviders and the
subsequent allocation of the corresponding resources to a
pod as can be seen here:
https://github.com/kubernetes/kubernetes/blob/v1.26.0/pkg/kubelet/cm/topologymanager/scope.go#L150
With regard to DRA, the passing of `cm.draManager` into
resourceAllocator seems redundant as no admission checks
(and allocation of resources handled by DRA) is taking place
in `Admit` method of resourceAllocator. DRA has a completely
different model to the rest of the resource managers where
pod is only scheduled on a node once resources are reserved
for it. Because of this, admission checks or waiting for
resources to be provisioned after the pod has been scheduled
on the node is not required.
Before making the above change, it was verified that DRA Manager
is instantiated in `NewContainerManager`:
https://github.com/kubernetes/kubernetes/blob/v1.26.0/pkg/kubelet/cm/container_manager_linux.go#L318
Signed-off-by: Swati Sehgal <swsehgal@redhat.com>
Since Topology manager is graduating to GA, we remove
internal configuration variable names with `Experimental`
prefix.
There is no expected change in behavior, only trival
variable renaming.
Signed-off-by: Swati Sehgal <swsehgal@redhat.com>
In case of node reboot/kubelet restart, the flow of events involves
obtaining the state from the checkpoint file followed by setting
the `healthDevices`/`unhealthyDevices` to its zero value. This is
done to allow the device plugin to re-register itself so that
capacity can be updated appropriately.
During the allocation phase, we need to check if the resources requested
by the pod have been registered AND healthy devices are present on
the node to be allocated.
Also we need to move this check above `needed==0` where needed is
required - devices allocated to the container (which is obtained from
the checkpoint file) because even in cases where no additional devices
have to be allocated (as they were pre-allocated), we still need to
make the devices that were previously allocated are healthy.
Signed-off-by: Swati Sehgal <swsehgal@redhat.com>
In order to implement the `full-pcpus-only` cpumanager policy option,
we leverage the implementation of the algorithm which picks CPUs.
By design, CPUs are taken from the biggest chunk available (socket
or NUMA zone) to physical cores, down to single cores.
Leveraging this, if the requested CPU count is a multiple of the SMT
level (commonly 2), we're guaranteed that only full physical cores
will be taken.
The hidden assumption here is this holds true by construction iff
the user reserved CPUs (if any) considering full physical CPUs.
IOW, if the user did intentionally or mistakely reserve single threads
which are no core siblings[1], then the simple check we implemented
is not sufficient.
A easy example can probably outline this better. With this setup:
cores: [(0, 4), (1, 5), (2, 6), (3, 8)] (in parens: thread siblings).
SMT level: 2 (each tuple is 2 elements)
Reserved CPUs: 0,1 (explicit pick using `--reserved-cpus`)
A container then requests 6 cpus. full-pcpus-only check: 6 % 2 == 0. Passed.
The CPU allocator will take first full cores, (2,6) and (3,8), and will
then pick the remaining single CPUs. The allocation will succeed, but
it's incorrect.
We can fix this case with a stricter precheck.
We need to additionally consider all the core siblings of the reserved
CPUs as unavailable when computing the free cpus, before to start the
actual allocation. Doing so, we fall back in the intended behavior, and
by construction all possible CPUs allocation whose number is multiple
of the SMT level are now correct again.
+++
[1] or thread siblings in the linux parlance, in any case:
hyperthread siblings of the same physical core
Signed-off-by: Francesco Romani <fromani@redhat.com>
