This Confluence has been LDAP enabled, if you are an ASF Committer, please use your LDAP Credentials to login. Any problems file an INFRA jira ticket please.

Child pages
  • Alex's Architecture Notes
Skip to end of metadata
Go to start of metadata

AgentManager and request dispatch

AgentManager interface is our abstraction layer that manages agent connection and message passing to agents. It's responsibility is as follows:

  • Provide abstraction layer such that all logic above does not need to know the particulars about how messages are serialize/deserialize and how messages are routed before reaching the agent.
  • Provide hook-ins to inform other pieces of code of agent connection and disconnect.
  • Provide hook-ins to inform outside code of Commands and Answers that has passed through the system.
  • Determine the nature of an agent disconnect.
  • Work with ResourceManager to determine whether an agent is allowed to connect
  • Scales well with the management servers in the cluster.

Let's break down some nomenclature:
A ServerResource in CloudStack is a translation layer between CloudStack operations and how to perform that operation on the physical resource that it interacts with. Examples of ServerResource are XenServer hypervisor, VmWare hypervisor, KVM hypervisor, F5, SRX, NetScaler, etc. The requirement for a ServerResource is for it to map a Command from CloudStack into operations performed on the physical resource without any database work. It is required that any ServerResource do not access the database.

An Agent in CloudStack is a container for an instance or instances of ServerResource. It's job is to serialize and deserialize the messages and make connections to the management server. You'll often see ServerResource and Agent usage being mixed up because Agent and instance of ServerResource are basically one to one today. However, that may not be true in the future. Just remember Agent is responsible for serialization and connection while ServerResource is responsible for execution. They are not the same.

Agents are broken down into different types.

  • Connected Agent is an agent that connects to the management server via the port 8250. KVM ServerResource is often contained within a Connected Agent.
  • Direct Agent is an agent that runs within the management server. XenServer ServerResource is often contained within a Direct Agent.
  • Forwarding Agent is an agent that routes messages between management servers

Command and Answer is our pattern for message requests and responses. Each Command should have a corresponding Answer. I have seen code that skips that but that's wrong and should be corrected.

Design of how to scale horizontally

The CloudStack management server have two sources of load. Obviously, one source is the number of requests it gets via the web services api. That's outside the scope of this email but we can talk about how that works in a separate email. The other source is the number of resources the management server cluster has to manage. Our objective is to make sure that we can simply add management servers to scale with the number of resources it manages. The following ensures that. - Agent Load Balancing: As management servers are started and stopped, agent load balancing rebalances the number of agent each management server handles without interrupting the message passing.

  • One ServerResource can only be connected to one management server at one time.
  • WebService API requests are always executed by one management server and if the ServerResource needed is on a different management server, AgentManager is responsible for routing the message to right management server.
  • Background tasks and monitoring processes on each management server only deals with the ServerResource connected to that management server. It can be notified of agent connections and disconnects by registering a Listener with the AgentManager.

Breakdown of how it performs each of these tasks.

Abstraction layer for serialization/deserialization

Messages are serialized/deserialized to json format by gson libraries. This code is encapsulated in and The actual content of the message depends on the Command and Answer that's being sent. You should look in package if you're interested in that. Note that the Commands and Answers are in the cloud-api.jar but and are in cloud-core.jar. We expect everyone who writes a resource to only depend on the cloud-api.jar. if you can't do that, then something is wrong with the design of the ServerResource.

Management Server clustering
  • AgentManagerImpl is responsible for implementing the process of message sending, receiving, Connected Agents, and Direct Agents.
  • ClusteredAgentManagerImpl extends AgentManagerImpl by adding features that only can happen in clustered situations: agent rebalancing and message routing between management servers.
  • AgentManager registers for events on management server entering and leaving the cluster with the ClusterManager. While ClusterManager is not technically part of AgentManager, it is important to describe its functionality. ClusterManager is responsible for notifying interested code that a management server node is up or down. It is does this by writing a heartbeat into the database and if the heartbeat has not been updated by a certain interval, it will inform all the interested parties. AgentManager is one of those interested. Upon notification that a management server node has left the system, AgentManager will pick up the agents that were connected to that management server.
Implementation of the different types of agent handling
  • Connected Agents are handled by, AgentHandler class embedded within, and The tcp connection itself is handled by a NioConnection class.
  • Direct Agents are handled by,
  • Forwarding Agents are handled by,
Notifying interested party of Agent Connection and Disconnect
  • Code interested in knowing if an agent is connected to the management server its running on can register a Listener with AgentManager. What it can listen for is in the interface.
Managing the life cycle of a ServerResource
  • That's implemented by ResourceManagerImpl. ResourceManager is responsible for the life cycle of a ServerResource and the groupings of ServerResource (Cluster, Pod, Zone). Now life cycle of a ServerResource does impact Agent connections. For example, a ServerResource that is in removed state should never have agents connected on its behalf. AgentManager, upon a connection, talks with the ResourceManager to determine if the connection is allowed.
Determining the cause of an agent disconnect

There may be many reasons why an agent disconnected. It could be because the physical resource is down. It could be because the physical resource is disconnected via tcp. Could be bugs within CloudStack code. has a disconnect method that all disconnect should go to. The job of this disconnect is to determine the appropriate action to take given what we know about the disconnects. VM HA is often triggered as part of this process.

AgentManager maintains an application ping with the agents. The ping interval is one minute but can be configured and the timeout is 2.5x the ping interval but can also be configured. If the application ping times out, AgentManager launches into an investigation process that tries to check if the physical resource is still alive. It does this by talking to a set of Investigators. Each Investigation can have three results: the resource is Up, the resource is Down, and I don't know. Upon receiving Up or Down, the AgentManager terminates the investigation and uses that state as the reason for the disconnect. Upon receiving "I don't know", AgentManager moves on to the next Investigator until it runs out of Investigators. It is crucial that the Investigator does not return a false positive or negative. For example, one particular Investigator pings the ip address of the physical resource. If it receives a ping response, then returns Up. However, if it doesn't receive a ping response, it should return "I don't know", because failure to respond to ping does not mean the physical resource is Down.

Resource Life Cycle

The life cycle of a particular resource really has the following states in

Creating - Being Created
Enabled - Enabled and can perform all commands
Disabled - Disabled and should not accept commands requiring new resource
PrepareForMaintenance - preparing to go into maintenance mode
ErrorInMaintenance - error during preparation. Someone needs to act on this.
Maintenance - Maintenance mode so no commands at all
Error - Something is wrong with the resource.
Unmanaged - leave the resource alone and don't allow any commands to be sent to it.

These states represent the administrator's intent for that resource. It really should be broken down in more detail. For example it probably should be more like the following but it is what it is for now.

  1. Enabled
  • Creating
  • Error
  1. Disabled
  2. Maintenance
  • Preparing
  • Error
  1. Unmanaged

For agent states, we have the following in

  • Creating - in the process of creating an attaché to talk to the agent
  • Connecting - in the process of connecting. During connection state, code via the Listener interface will send commands to the resource to ensure the resource is in a good enough state to handle commands from the management server.
  • Up - up and running
  • Down - the physical resource is down!!! Note that this can only happen if an investigator returned and said the resource is down.
  • Disconnected - agent is disconnected
  • Alert - agent has been disconnected for more than 30 minutes (configurable value)
  • Removed - attaché is removed.
  • Error - Error during connection
  • Rebalancing - attaché is being rebalanced