The scalable data service has the potential for active instances on multiple nodes. Scalable services use two resource groups: a scalable resource group to contain the application resources and a failover resource group to contain the network resources (shared addresses) on which the scalable service depends. The scalable resource group can be online on multiple nodes, so multiple instances of the service can be running at once. The failover resource group that hosts the shared address is online on only one node at a time. All nodes hosting a scalable service use the same shared address to host the service.
Service requests come into the cluster through a single network interface (the global interface) and are distributed to the nodes based on one of several predefined algorithms set by the load-balancing policy. The cluster can use the load-balancing policy to balance the service load between several nodes. Note that there can be multiple global interfaces on different nodes hosting other shared addresses.
For scalable services, application instances run on several nodes simultaneously. If the node that hosts the global interface fails, the global interface fails over to another node. If an application instance running fails, the instance attempts to restart on the same node.
If an application instance cannot be restarted on the same node, and another unused node is configured to run the service, the service fails over to the unused node. Otherwise, it continues to run on the remaining nodes, possibly causing a degradation of service throughput.
TCP state for each application instance is kept on the node with the instance, not on the global interface node. Therefore, failure of the global interface node does not affect the connection.
The figure Failover and Scalable Resource Group Example shows an example of failover and a scalable resource group and the dependencies that exist between them for scalable services. This example shows three resource groups. The failover resource group contains application resources for highly available DNS, and network resources used by both highly available DNS and highly available Apache Web Server. The scalable resource groups contain only application instances of the Apache Web Server. Note that resource group dependencies exist between the scalable and failover resource groups (solid lines) and that all of the Apache application resources are dependent on the network resource schost-2, which is a shared address (dashed lines).
The primary goal of cluster networking is to provide scalability for data services. Scalability means that as the load offered to a service increases, a data service can maintain a constant response time in the face of this increased workload as new nodes are added to the cluster and new server instances are run. We call such a service a scalable data service. A good example of a scalable data service is a web service. Typically, a scalable data service is composed of several instances, each of which runs on different nodes of the cluster. Together these instances behave as a single service from the standpoint of a remote client of that service and implement the functionality of the service. We might, for example, have a scalable web service made up of several httpd daemons running on different nodes. Any httpd daemon may serve a client request. The daemon that serves the request depends on a load-balancing policy. The reply to the client appears to come from the service, not the particular daemon that serviced the request, thus preserving the single service appearance.
A scalable service is composed of:
Networking infrastructure support for scalable services
Support for networking and data services (using the Resource Group Manager)
The following figure depicts the scalable service architecture.
The nodes that are not hosting the global interface (proxy nodes) have the shared address hosted on their loopback interfaces. Packets coming into the global interface are distributed to other cluster nodes based on configurable load-balancing policies. The possible load-balancing policies are described next.
Load balancing improves performance of the scalable service, both in response time and in throughput.
There are two classes of scalable data services: pure and sticky. A pure service is one where any instance of it can respond to client requests. A sticky service is one where a client sends requests to the same instance. Those requests are not redirected to other instances.
A pure service uses a weighted load-balancing policy. Under this load-balancing policy, client requests are by default uniformly distributed over the server instances in the cluster. For example, in a three-node cluster, let us suppose that each node has the weight of 1. Each node will service 1/3 of the requests from any client on behalf of that service. Weights can be changed at any time by the administrator through the scrgadm(1M) command interface or through the SunPlex Manager GUI.
A sticky service has two flavors, ordinary sticky and wildcard sticky. Sticky services allow concurrent application-level sessions over multiple TCP connections to share in-state memory (application session state).
Ordinary sticky services permit a client to share state between multiple concurrent TCP connections. The client is said to be “sticky” with respect to that server instance listening on a single port. The client is guaranteed that all of his requests go to the same server instance, provided that instance remains up and accessible and the load balancing policy is not changed while the service is online.
For example, a web browser on the client connects to a shared IP address on port 80 using three different TCP connections, but the connections are exchanging cached session information between them at the service.
A generalization of a sticky policy extends to multiple scalable services exchanging session information behind the scenes at the same instance. When these services exchange session information behind the scenes at the same instance, the client is said to be“sticky” with respect to multiple server instances on the same node listening on different ports.
For example, a customer on an e-commerce site fills his shopping cart with items using ordinary HTTP on port 80, but switches to SSL on port 443 to send secure data in order to pay by credit card for the items in the cart.
Wildcard sticky services use dynamically assigned port numbers, but still expect client requests to go to the same node. The client is “sticky wildcard” over ports with respect to the same IP address.
A good example of this policy is passive mode FTP. A client connects to an FTP server on port 21 and is then informed by the server to connect back to a listener port server in the dynamic port range. All requests for this IP address are forwarded to the same node that the server informed the client through the control information.
Note that for each of these sticky policies the weighted load-balancing policy is in effect by default, thus, a client's initial request is directed to the instance dictated by the load balancer. After the client has established an affinity for the node where the instance is running, then future requests are directed to that instance as long as the node is accessible and the load balancing policy is not changed.
Additional details of the specific load balancing policies are discussed below.
Weighted. The load is distributed among various nodes according to specified weight values. This policy is set using the LB_WEIGHTED value for the Load_balancing_weights property. If a weight for a node is not explicitly set, the weight for that node defaults to one.
The weighted policy redirects a certain percentage of the traffic from clients to a particular node. Given X=weight and A=the total weights of all active nodes, an active node can expect approximately X/A of the total new connections to be directed to the active node, when the total number of connections is large enough. This policy does not address individual requests.
Note that this policy is not round robin. A round-robin policy would always cause each request from a client to go to a different node: the first request to node 1, the second request to node 2, and so on.
Sticky. In this policy, the set of ports is known at the time the application resources are configured. This policy is set using the LB_STICKY value for the Load_balancing_policy resource property.
Sticky-wildcard. This policy is a superset of the ordinary“sticky” policy. For a scalable service identified by the IP address, ports are assigned by the server (and are not known in advance). The ports might change. This policy is set using the LB_STICKY_WILD value for the Load_balancing_policy resource property.