Architecture¶
Overview¶
CRI Resource Manager (CRI-RM) is a pluggable add-on for controlling how much and which resources are assigned to containers in a Kubernetes cluster. It’s an add-on because you install it in addition to the normal selection of your components. It’s pluggable since you inject it on the signaling path between two existing components with the rest of the cluster unaware of its presence.
CRI-RM plugs in between kubelet and CRI, the Kubernetes node agent and the container runtime implementation. It intercepts CRI protocol requests from the kubelet acting as a non-transparent proxy towards the runtime. Proxying by CRI-RM is non-transparent in nature because it usually alters intercepted protocol messages before forwarding them.
CRI-RM keeps track of the states of all containers running on a Kubernetes node. Whenever it intercepts a CRI request which results in changes to the resource allocation of any container (container creation, deletion or resource assignment update request), CRI-RM runs one of its built-in policy algorithms. This policy makes a decision about how the assignment of resources should be updated and, eventually, the intercepted request gets modified according to this decision. The policy can make changes to any container in the system, not just the one associated with the intercepted CRI request. Therefore it does not operate directly on CRI requests. Instead CRI-RM’s internal state tracking cache provides an abstraction for modifying containers and the policy uses this abstraction for recording its decisions.
In addition to policies, CRI-RM has a number of built-in resource controllers. These are used to put policy decisions, in practice pending changes made to containers by a policy, into effect. A special in-band CRI controller is used to control all resources which are controllable via the CRI runtime. This controller handles the practical details of updating the intercepted CRI request and generating any additional unsolicited update requests for other existing containers updated by the policy decision. Additional out-of-band controllers exist to exercise control over resources which the current CRI runtimes are unable to handle.
To tell which containers need to be handed off to various controllers for updating, CRI-RM uses the internal state tracking cache’s ability to tell which containers have pending unenforced changes and to which controllers’ domain these changes belong to. The CRI controller currently handles CPU and memory resources, including huge pages. The level of control covers per container CPU sets, CFS parametrization, memory limits, OOM score adjustment and pinning to memory controllers. The two existing out-of-band controllers, RDT and Block I/O, handle last level cache and memory bandwidth allocation, and the arbitration of Block I/O bandwidth respectively.
Many of the details of how CRI-RM operates is configurable. These include, for instance, which policy is active within CRI-RM, configuration of the resource assignment algorithm for the active policy, and configuration for the various resource controllers. Alhough CRI-RM can be configured using a configuration file present on the node running CRI-RM, the preferred way for configuring all CRI-RM instances in a cluster is to use Kubernetes ConfigMaps and the CRI-RM Node Agent.
Components¶
Node Agent¶
The node agent is a component external to CRI-RM itself. All interactions by CRI-RM with the Kubernetes Control Plane go through the node agent with the node agent performing any direct interactions on behalf of CRI-RM.
The node agent communicates with CRI-RM using two gRPC interfaces. The config interface is used to
push updated external configuration data to CRI-RM
push adjustments to container resource assignments to CRI-RM
The cluster interface implements the necessary low-level plumbing for the agent interface CRI-RM internally exposes for its policies and other components. This interface in turn implements
updating resource capacity of the node
getting, setting or removing labels on the node
getting, setting or removing annotations on the node
getting, setting or removing taints on the node
The config interface is defined and has its gRPC server running in CRI-RM. The agent acts as a gRPC client for this interface. The low-level cluster interface is defined and has its gRPC server running in the agent, with the convenience layer defined in CRI-RM. CRI-RM acts as a gRPC client for the low-level plumbing interface.
Additionally, the stock node agent that comes with CRI-RM implements schemes for
configuration management for all CRI-RM instances
management of dynamic adjustments to container resource assignments
Resource Manager¶
CRI-RM implements a request processing and an event processing pipeline. The request processing pipeline takes care of proxying CRI requests and responses between CRI clients and the CRI runtime. The event processing pipeline processes a set of other events which are not directly related to or the result of CRI requests. These events are typically internally generated within CRI-RM. They can be the result of changes in the state of some containers or the utilisation of a shared system resource, which potentially could warrant an attempt to rebalance the distribution of resources among containers to bring the system closer to an optimal state. Some events can also be generated by policies.
The Resource Manager component of CRI-RM implements the basic control flow of both of these processing pipelines. It passes control to all the necessary sub-components of CRI-RM at the various phases of processing a request or an event. Additionally it serializes the processing of these, making sure there is at most one (intercepted) request or event being processed at any point in time.
The high-level control flow of the request processing pipeline is the following:
A. If the request does not need policying, let it bypass the processing pipeline: hand it off for logging, then relay it to the server and the corresponding response back to the client.
B. If the request needs to be intercepted for policying
take processing pipeline serialization lock
look up/create cache objects (pod/container) for the request
if the request has no resource allocation consequences, do proxying (step 6)
otherwise invoke the policy layer for resource allocation:
pass it on to the configured active policy, which will
allocate resources for the container
update the assignments for the container in the cache
update any other containers affected by the allocation in the cache
invoke the controller layer for post-policy processing, which will
collect controllers with pending changes in their domain of control
for each invoke the post-policy processing function corresponding to the request
clear pending markers for the controllers
proxy the request
relay the request to the server
send update requests for any additional affected containers
update the cache if/as necessary based on the response
relay the response back to the client
release processing pipeline serialization lock
The high-level control flow of the event processing pipeline is one of the following, based on the event type:
for policy-specific events:
take processing pipeline lock
call policy event handler
invoke the controller layer for post-policy processing (same as step 5 for requests)
release pipeline lock
for metrics events:
perform collection/processing/correlation
take processing pipeline lock
update cache objects as/if necessary
request rebalancing as/if necessary
release pipeline lock
for rebalance events:
take processing pipeline lock
invoke policy layer for rebalancing
invoke the controller layer for post-policy processing (same as step 5 for requests)
release pipeline lock
Cache¶
The cache is a shared internal storage within CRI-RM. It tracks the runtime state of pods and containers known to CRI-RM, as well as the state of CRI-RM itself, including the active configuration and the state of the active policy. The cache is saved to permanent storage in the filesystem and used to restore the runtime state of CRI-RM across restarts.
The cache provides functions for querying and updating the state of pods and containers. This is the mechanism used by the active policy to make resource assignment decisions. The policy simply updates the state of the affected containers in the cache according to the decisions.
The cache’s ability to associate and track changes to containers with resource domains is used to enforce policy decisions. The generic controller layer first queries which containers have pending changes, then invokes each controller for each container. The controllers use the querying functions provided by the cache to decide if anything in their resource/control domain needs to be changed and then act accordingly.
Access to the cache needs to be serialized. However, this serialization is not provided by the cache itself. Instead, it assumes callers to make sure proper protection is in place against concurrent read-write access. The request and event processing pipelines in the resource manager use a lock to serialize request and event processing and consequently access to the cache.
If a policy needs to do processing unsolicited by the resource manager, IOW processing other than handling the internal policy backend API calls from the resource manager, then it should inject a policy event into the resource managers event loop. This will cause a callback from the resource manager to the policy’s event handler with the injected event as an argument and with the cache properly locked.
Generic Policy Layer¶
The generic policy layer defines the abstract interface the rest of CRI-RM uses to interact with policy implementations and takes care of the details of activating and dispatching calls through to the configured active policy.
Generic Resource Controller Layer¶
The generic resource controller layer defines the abstract interface the rest of CRI-RM uses to interact with resource controller implementations and takes care of the details of dispatching calls to the controller implementations for post-policy enforcment of decisions.
Metrics Collector¶
The metrics collector gathers a set of runtime metrics about the containers running on the node. CRI-RM can be configured to periodically evaluate this collected data to determine how optimal the current assignment of container resources is, and to attempt a rebalancing/reallocation if it is deemed both possible and necessary.
Policy Implementations¶
None¶
An empty policy which does no policy decisions. It is included merely for the sake of completeness, analoguous to the none policy of the CPU Manager in kubelet.
Static Pools¶
A backward-compatible reimplementation of CMK for CRI-RM with a few extra features.
Static¶
Part of the code from the static policy of CPU Manager in kubelet, brutally hacked to work within CRI-RM. Serves merely as a proof-of-concept that the current policies of kubelet can be implemented in CRI-RM.
Static Plus¶
A fairly simplistic policy similar in spirit to the static policy of CPU Manager in kubelet, with a few extra features.
Resource Controller Implementations¶
RDT¶
A resource controller implementation responsible for the practical details of associating a container with RDT classes. This class effectively determines how much last level cache and memory bandwidth will be available for the container. This controller uses the resctrl pseudo filesystem of the Linux kernel for control.
Block I/O¶
A resource controller implementation responsible for the practical details of associating a container with a Block I/O class. This class effectively determines how much Block I/O bandwidth will be available for the container. This controller uses the blkio cgroup controller and the cgroupfs pseudo- filesystem of the Linux kernel for control.