1 Introduction to Oracle Database QoS Management

You can use server pools to create groups of servers within a cluster to provide workload isolation. A server can only belong to one server pool at any time. You can create an Oracle Database in a single server pool, or across multiple server pools. Oracle Database QoS Management can make recommendations to move a server from one server pool to another based on the measured and projected demand. Oracle Database QoS Management can also relocate servers to satisfy the Performance Objectives currently in effect.

When multiple database instances share a single server, they must share its CPU, memory, and I/O bandwidth. Instance Caging limits the amount of CPU an Oracle database instance consumes by using the Oracle Database Resource Manager and the CPU_COUNT database initialization parameter. When using Oracle Database QoS Management, the sum of the values for CPU_COUNT for all instances of the server must be less than or equal to the total number of physical CPUs. Also, each CPU partition, or slice, must be uniform in thickness (number of CPUs) for each instance of a database within a server pool. These requirements help to ensure predictable and isolated performance of each database.

Figure 1-1 Instance Caging and CPU Slices

Description of "Figure 1-1 Instance Caging and CPU Slices"

When you implement instance caging, Oracle Database QoS Management can provide recommendations to reallocate CPU resources from one slice to another slice within the same server pool. If you choose to implement the recommendation to modify the instance caging settings, then Oracle Database QoS Management modifies the CPU_COUNT parameter uniformly for all the database instances running on servers in the server pool.

Modifying the CPU_COUNT parameter and configuring Oracle Database QoS Management so that a Resource Plan is activated enables the Instance Caging feature. When you use Instance Caging to constrain CPU usage for an instance, that instance could become CPU-bound. This is when the Resource Manager begins to do its work, allocating CPU shares among the various database sessions according to the active resource plan.

In an Oracle RAC cluster, Oracle Database QoS Management monitors the server pools and nodes on which the database services are offered. A service can run in only one server pool. If the database spans multiple server pools, then you must create multiple services to access the instances in all server pools.

Workload is monitored for clients and applications that connect to the database using database services that are managed by Oracle Clusterware. The connections must use Java Database Connectivity (JDBC) (thick or thin), or Oracle Call Interface (OCI). Connections should use services with its run-time goal for the load balancing advisory set to SERVICE_TIME, and the connection load balancing goal set to LONG. For example:

srvctl modify service -db db_name -service service_name -rlbgoal SERVICE_TIME 
-clbgoal LONG -cardinality UNIFORM

You must define the cardinality of database services as follows:

• If the server pool that the service runs in has a maximum size greater than 1 (or UNLIMITED), then set the cardinality of the service to UNIFORM.

• If the server pool that the service runs in has a maximum size of 1, then set the cardinality of the service to SINGLETON.

The classification of work requests applies a user-defined name (tag) that identifies the Performance Class to which the work request belongs. All work requests that are grouped into a particular Performance Class have the same performance objectives. In effect, the tag connects the work request to the Performance Objective for the associated Performance Class. Oracle Database QoS Management assigns tags to each work request so that every component of the system can take measurements, and provide data to Oracle Database QoS Management for evaluation against the applicable Performance Objectives.

Classification occurs wherever new work enters the system. When a work request arrives at a server, the work request is checked for a tag. If the work request has a tag, then the server concludes that this work request has already been classified, and the tag is not changed. If the work request does not include a tag, then the classifiers are checked, and a tag for the matching Performance Class is attached to the work request.

To illustrate how work requests are classified, consider an application that connects to an Oracle RAC database. The application uses the database service sales. The Oracle Database QoS Management administrator specified during the initial configuration of Oracle Database QoS Management that the sales_pc Performance Class should contain work requests that use the sales service. When a connection request is received by the database, Oracle Database QoS Management checks for a tag. If a tag is not found, then Oracle Database QoS Management compares the information in the connection request with the classifiers specified for each Performance Class, in the order specified in the Performance Policy. If the connection request being classified is using the sales service, then when the classifiers in the sales_pc Performance Class are compared to the connection request information, a match is found, and the database work request is assigned a tag for the sales_pc Performance Class.

A single application can support work requests of many types, with a range of performance characteristics. By extending and refining the default classification rules, the Oracle Database QoS Management administrator can write multiple Performance Objectives for a single application. For example, the administrator may decide that a web-based application should have separate Performance Objectives for work requests related to logging in, browsing, searching, and purchasing.

Oracle Database QoS Management supports user-defined combinations of connection parameters to map Performance Classes to the actual workloads running in the database. These connection parameters belong to two general classes, and can be combined to create fine-grained Boolean expressions:

• Configuration Parameters—The supported configuration parameters are SERVICE_NAME and USERNAME. Each classifier in a Performance Class must specify the name of a database service. Additional granularity can be achieved by identifying the name of the user that is making the database connection from either a client or the middle tier. The advantage of using these classifiers is that they do not require application code changes to associate different workloads with separate Performance Classes.

• Application Parameters—The supported application parameters are MODULE, ACTION, and PROGRAM. These are optional parameters. The values for MODULE and ACTION must be set within the application. Depending on the type of application, you can set these parameters as follows:

  • OCI—Use OCI_ATTR_MODULE and OCI_ATTR_ACTION.

  • Oracle Data Provider for .NET (ODP.NET)—Specify the ModuleName and ActionName properties on the OracleConnection object.

  • JDBC—Set MODULE and ACTION in SYS_CONTEXT.

The PROGRAM parameter is set or derived differently for each database driver and platform. Consult the appropriate Oracle Database developer's guide for further details and examples.

To manage the workload for an application, the application code makes database connections using a particular service. To provide more precise control over the workload generated by various parts of the application, you can create additional Performance Classes, and use classifiers that include PROGRAM, MODULE, or ACTION in addition to the service or user name. For example, you could specify that all connections to your cluster that use the sales service belong to the sales_pc Performance Class, but connections that use the sales service and have a user name of APPADMIN belong to sales_admin Performance Class.

The Performance Classes in use at a particular data center are expected to change over time. For example, you may need to modify the Performance Objectives for one part of your application. In this case you would create a new Performance Class with additional classifiers to identify the target work requests, and update your Performance Policy to add a new Performance Objective for this Performance Class. In other words, you replace a single Performance Objective with one or more finer-grained Performance Objectives, and divide the work requests for one Performance Class into multiple Performance Classes.

Application developers can suggest which Performance Classes to use. Specifically, an application developer can suggest ways to identify different application workloads, and you can use these suggestions to create classifiers for Performance Classes so that each type of work request is managed separately.

You can create additional Performance Classes to specify acceptable response times for different application workloads. For example, a Performance Objective may indicate that a work request performing the checkout action for the sales_pc_checkout Performance Class should not take more than one millisecond to complete, and a work request performing the browse action for the sales_pc_browse Performance Class can take 100 milliseconds second to complete.

You create Performance Objectives for each Performance Class to specify the target performance level for all work requests that are assigned to each Performance Class. A Performance Objective specifies both a business requirement (the target performance level) and the work to which that Performance Objective applies (the Performance Class). For example, a Performance Objective could specify that work requests in the hr_pc Performance Class should have an average response time of less than 0.2 seconds.

Performance Objectives are specified with Performance Policies. Each Performance Policy includes a Performance Objective for each and every Performance Class, unless the Performance Class is marked Measure-Only. In this release, Oracle Database QoS Management supports only one type of Performance Objective, average response time.

The response time for a workload is based upon database client requests. Response time measures the time from when the cluster receives the request over the network to the time the request leaves the cluster. Response time does not include the time required to send the information over the network to or from the client. The response time for all database client requests in a Performance Class is averaged and presented as average response time, measured as database requests for second.

A server pool directive override sets the availability properties of Min, Max, and Importance for a server pool when the Performance Policy is in effect.

A Performance Policy can also include a set of server pool directive overrides. Server pool directive overrides serve as constraints on the allocation changes that Oracle Database QoS Management recommends, because the server pool directive overrides are honored during the activation period of the Performance Policy. For example, Oracle Database QoS Management never recommends moving a server out of a server pool if doing so results in the server pool having less than its specified minimum number of servers.

You could create Performance Policies for your system to manage workload based on the time of year or time of day, as shown in Figure 1-5. Under normal conditions, these Performance Policies keep your database workload running at a steady rate. If the workload requests for a database increase suddenly, then a particular server pool may require additional resources beyond what is specified by the Performance Policy.

Figure 1-5 Baseline Resource Management by Performance Policy

Description of "Figure 1-5 Baseline Resource Management by Performance Policy"

For example, assume your business takes orders over the telephone, and creates orders using a sales application. Your telephone sales department is only open during regular business hours, but customers can also place orders themselves over the Internet. During the day, more orders are placed so the sales applications need more resources to handle the workload. This configuration is managed by creating the Business Hours Performance Policy, and specifying that the Back Office server pool can have a maximum of two servers, enabling Oracle Database QoS Management to move servers to the Online server pool, as needed. After the telephone sales department closes, the workload for the sales applications decreases. To manage this configuration you create the After Hours Performance Policy and specify that the Back Office server pool can have a maximum of four servers, enabling your internal applications to acquire the additional resources that they need to complete their workloads before the next business day.

In this scenario, the Business Hours and After Hours Performance Policies can contain server pool directive overrides. When a Performance Policy contains a server pool directive override, the current settings of Max, Min, and Importance for the specified server pool are overridden while that Performance Policy is in effect. This enables additional servers to be placed in the Sales server pool to give the online sales applications the resources they need and to limit the resources used by the Back Office server pool, so that its workload does not interfere with the Sales workload.

Within a Performance Policy, you can also assign a level of business criticalness (a rank) to each Performance Class to give priority to meeting the Performance Objectives for a more critical Performance Class over a less critical one. When there are not enough resources available to meet all the Performance Objectives for all Performance Classes at the same time, the Performance Objectives for the more critical Performance Classes must be met at the expense of the less critical Performance Objectives. The Performance Policy specifies the business criticalness of each Performance Class, which can be Highest, High, Medium, Low, or Lowest.

For example, using the Performance Policies illustrated in Figure 1-5, when the Business Hours Performance Policy is in effect, the sales applications, which access the Online server pool, have the highest rank. If there are not enough resources available to meet the Performance Objectives of all the Performance Classes, then the applications that use the Online server pool will get priority access to any available resources, even if the applications using the Back Office server pool are not meeting their Performance Objectives.

You can have multiple Performance Classes at the same rank. If Oracle Database QoS Management detects more than one Performance Class not meeting its Performance Objective and the Performance Classes are assigned the same rank in the active Performance Policy, then Oracle Database QoS Management recommends a change to give the Performance Class closest to meeting its Performance Objective more resources. After implementing the recommended action, when the Performance Class is no longer below its target performance level, Oracle Database QoS Management performs a new evaluation of the system performance.

If a user-created Oracle Database QoS Management policy is active, then the ranked order of the performance classes determines the order in which the associated Oracle RAC databases start or request real-time LMS process slots. Using the performance class rankings ensures that mission critical databases running in a consolidated environment have their LMS processes run in real-time, thus eliminating a resource bottleneck within inter-node communication. Because the Oracle Database QoS Management policy specifies the rank of each workload, using the value of Max(Ranks) for each database provides a consistent expression of the expressed business criticality of each database.

When you use Oracle Database QoS Management, your system is continuously monitored in an iterative process to see if the Performance Objectives in the active Performance Policy are being met. Performance data are sent to Oracle Enterprise Manager for display in the Oracle Database QoS Management Dashboard (the Dashboard) and Performance History pages.

When one or more Performance Objectives are not being met, after evaluating the performance of your system, Oracle Database QoS Management seeks to improve the performance of a single Performance Objective: usually the highest ranked Performance Objective that is currently not being satisfied. If all Performance Objectives are satisfied with capacity to spare for both the current and projected workload, then Oracle Database QoS Management signals "No action required: all Performance Objectives are being met."

If Performance Objectives are not being met for a Performance Class, then Oracle Database QoS Management issues recommendations to rebalance the use of resources to alleviate bottlenecks. Oracle Database QoS Management evaluates several possible solutions and then chooses the solution that:

• Offers the best overall system improvement

• Causes the least system disruption

• Helps the highest ranked violating performance class

The types of recommendations that Oracle Database QoS Management can make are:

• Promoting and Demoting Consumer Groups

• Moving Servers Between Server Pools

• Modifying the CPU Count

When trying to relieve a resource bottleneck for a particular Performance Class, Oracle Database QoS Management recommends adding more of the resource (such as CPU time) for that Performance Class or making the resource available more quickly to work requests in the Performance Class. The recommendations take the form of promoting the target Performance Class to a higher Consumer Group, demoting competing Performance Classes within the resource plan, adjusting CPU resources shared between different slices in a server pool, or moving servers between server pools.

Implementing a recommended action makes the resource less available to other Performance Classes. When generating recommendations, Oracle Database QoS Management evaluates the impact to system performance as a whole. If a possible recommendation for changing the allocation of resources provides a small improvement in the response time of one Performance Class, but results in a large decrease in the response time of another Performance Class, then Oracle Database QoS Management reports that the performance gain is too small, and the change is not recommended.

Oracle Database QoS Management can issue recommendations that involve a negative impact to the performance of a Performance Class if:

• The negative impact on the Performance Class from which the resource is taken is projected not to cause a Performance Objective violation and a positive impact is projected for the Performance Class that gets better access to resources

• The Performance Class from which the resource is taken is lower ranked, and thus less critical to your business, than the Performance Class being helped

If the resource bottleneck can be resolved in multiple ways, then Oracle Database QoS Management recommends an action that is projected to improve the performance of the highest ranked Performance Class that is violating its objective. You can also view the alternative recommendations generated by Oracle Database QoS Management and see whether the action was recommended for implementation. For example, one possible solution to resolving a bottleneck on the CPU resource is to demote the Consumer Group associated with the Performance Class that is using the CPU the most. By limiting access to the CPU for the work requests in this Performance Class, the work requests in the other Performances Classes for that database get a larger share of the CPU time. However, Oracle Database QoS Management may decide not to recommend this action because the gain in response time for the target Performance Class is too small.

The analysis data for a recommendation include the projected change in response time for each Performance Class, the projected change in the Performance Satisfaction Metric (PSM) for each Performance Class, and the reason this action is chosen among other alternative actions, as shown in Figure 1-7. In this example, if you implement the recommended action, then Oracle Database QoS Management predicts that the sales cart Performance Class, which has the highest ranking, will have an improvement in response time from 0.00510 seconds for database requests to 0.00426 seconds, which equates to an 11.6% gain in its PSM. The other Performance Classes are not effected by the change because they use a different server pool.

Figure 1-7 Example of the Analysis for a Recommended Action

Description of "Figure 1-7 Example of the Analysis for a Recommended Action"

Oracle Database QoS Management does not implement the recommendations automatically, however, you can configure Enterprise Manager to generate alerts based upon the duration that a Performance Class has not been meeting its objective. After the Oracle Database QoS Management administrator implements a recommendation, the system performance is reevaluated for the specified settling time before any new recommendations are made.

Consider a system that has two servers in an Online server pool, and two servers in a Back Office server pool. The Online server pool hosts two workloads: the sales_pc Performance Class and the sales_cart Performance Class. The minimum size of the Online server pool is two. The Back Office server pool hosts two internal applications: a human resources (HR) application and an enterprise resource planning (ERP) application. The Back Office server pool has a minimum size of one. The sales_cart Performance Class has the highest rank and the erp_pc Performance Class has the lowest rank. The sales_pc Performance Class is ranked higher than the hr_pc Performance Class.

In this scenario, if the sales_pc workload surges, causing contention for resources and causing the sales_cart Performance Class to violate its Performance Objective, then this could lead to a service-level agreement (SLA) violation for the OLTP application. Oracle Database QoS Management issues a recommendation to increase access to the CPU for the sales_cart Performance Class at the expense of the sales_pc workload, because the sales_cart Performance Class is of a higher rank; a higher rank indicates that satisfying the Performance Objective for the sales_cart Performance Class is more important than satisfying the Performance Objective for the sales_pc Performance Class.

If, after you implement the recommendation, the sales_cart and sales_pc Performance Classes are still not satisfying their Performance Objectives, then Oracle Database QoS Management issues a recommendation to increase the number of servers in the Online server pool by moving a server from the Back Office server pool, or a server pool that hosts less critical workloads or workloads with more headroom. In this scenario, a server can be moved from the Back Office server pool, because the Back Office server pool is currently above its minimum size of one. If the Back Office server pool had a minimum size of two, then Oracle Database QoS Management would have to find an available server in a different server pool; Oracle Database QoS Management does not recommend to move a server from a server pool if doing so will cause a server pool to drop below its minimum size.

If you implement the recommended action, and your applications use Cluster Managed Services and Client Run-time Load Balancing, then the application users should not see a service disruption due to this reallocation. The services are shut down transactionally on the server being moved. After the server has been added to the stressed server pool, all database instances and their offered services are started on the reallocated server. At this point, sessions start to gradually switch to using the new server in the server pool, relieving the bottleneck.

Using the same scenario, if the sales_pc Performance Class and hr_pc Performance Class both require additional servers to meet their Performance Objectives, then Oracle Database QoS Management first issues recommendations to improve the performance of the sales_pc Performance Class, because the sales_pc Performance Class is ranked higher than the hr_pc Performance Class. When the sales_pc Performance Class is satisfying its Performance Objectives, then Oracle Database QoS Management makes recommendations to improve the performance of the hr_pc Performance Class.

You create database services to provide a mechanism for grouping related work requests. An application connects to the cluster databases using database services. A user-initiated query against the database could use a different service than a web-based application. Different services can represent different types of work requests. Each call or request made to the Oracle RAC database is a work request.

You can also use database services to manage and measure database workloads. To manage the resources used by a service, some services may be deployed on several Oracle RAC instances concurrently, whereas others may be deployed on only a single instance to isolate the workload that uses that service.

In an Oracle RAC cluster, Oracle Database QoS Management monitors the server pools and its nodes, on which the database services are offered. Services are created by the database administrator for a database. For a policy-managed database, the service runs on all servers in the specified server pool. If a singleton service is required due to the inability of the application to scale horizontally, then the service can be restricted to run in a server pool that has a minimum and maximum size of one.

To use Oracle Database QoS Management for managing performance, create one or more policy-managed databases that run in server pools. If you have administrator-managed databases, then the database instances are placed in the Generic server pool and Oracle Database QoS Management can only monitor these databases.

When you first configure Oracle Database QoS Management, a default Performance Policy is created for each service that is discovered on the server pools being monitored. The name of these default Performance Classes are service_name_pc. The workload you want to monitor and manage the resource for must use a database service to connect to the database.

Oracle Database Resource Manager (Resource Manager) is an example of a resource allocation mechanism; Resource Manager can allocate CPU shares among a collection of resource consumer groups based on a resource plan specified by an administrator. A resource plan allocates the percentage of opportunities to run on the CPU.

Oracle Database QoS Management does not adjust existing Resource Manager plans; Oracle Database QoS Management activates a resource plan named APPQOS_PLAN, which is a complex, multilevel resource plan. Oracle Database QoS Management also creates consumer groups that represent Performance Classes and resource plan directives for each consumer group.

When you implement an Oracle Database QoS Management recommendation to promote or demote a consumer group for a Performance Class, Oracle Database QoS Management makes the recommended changes to the mapping of the Performance Class to the CPU shares specified in the APPQOS_PLAN resource plan. By altering the consumer group, the Performance Class that is currently not meeting its Performance Objective is given more access to the CPU resource.

By default, the APPQOS_PLAN is replaced during the Oracle Scheduler maintenance window. Oracle recommends that you use the APPQOS_PLAN as the plan during those daily windows because this resource plan incorporates the consumer groups from the DEFAULT_MAINTENANCE_PLAN plan. You can force the use of APPQOS_PLAN by running the following commands in SQL*Plus:

BEGIN
 DBMS_SCHEDULER.DISABLE(name=>'"SYS"."MONDAY_WINDOW"');
END;
/

BEGIN
 DBMS_SCHEDULER.SET_ATTRIBUTE(name=>'"SYS"."MONDAY_WINDOW"',
   attribute=>'RESOURCE_PLAN',value=>'APPQOS_PLAN');
END;
/

BEGIN
 DBMS_SCHEDULER.ENABLE(name=>'"SYS"."MONDAY_WINDOW"');
END;
/

Repeat these commands for every weekday, for example, TUESDAY_WINDOW.

You must have Oracle Clusterware installed and configured before you can use Oracle Database QoS Management. The cluster administrator should create server pools to be used to deploy policy-managed Oracle RAC databases. Administrator-managed Oracle RAC databases use only the Generic server pool.

When you first configure Oracle Database QoS Management and create the initial Policy Set, you specify which server pools should be managed by Oracle Database QoS Management and which should only be monitored. If you select a server pool to be managed by Oracle Database QoS Management, then Oracle Database QoS Management monitors the resources used by all the Performance Classes that run in that server pool. If a Performance Class is not satisfying its Performance Objective, then Oracle Database QoS Management can recommend moving servers between server pools to provide additional resources where needed.

Oracle Database QoS Management uses Oracle Cluster Health Monitor (CHM) to collect memory metric data for the servers in the cluster.

Run-time connection load balancing enables Oracle Clients to provide intelligent allocations of connections in the connection pool when applications request a connection to complete some work; the decision of which instance to route a new connection to is based on the current level of performance provided by the database instances.

Applications that use resources managed by Oracle Database QoS Management can also benefit from connection load balancing and transparent application failover (TAF). Connection load balancing enables you to spread user connections across all of the instances that are supporting a service. For each service, you can define the method you want the listener to use for load balancing by setting the connection load balancing goal, using the appropriate SRVCTL command with the -clbgoal option. You can also specify a single TAF policy for all users of a service using SRVCTL with the options -failovermethod,-failovertype, and so on.