In IBM mainframes, Workload Manager (WLM) is a base component of MVS/ESA mainframe operating system, and its successors up to and including z/OS. It controls the access to system resources for the work executing on z/OS based on administrator-defined goals. Workload Manager components also exist for other operating systems. For example, an IBM Workload Manager is also a software product for AIX operating system.

Workload Manager

On a mainframe computer many different applications execute at the same time. The expectations for executing work are consistent execution times and predictable access to databases. On z/OS the Workload Manager (WLM) component fulfills these needs by controlling work's access to system resources based on external specifications by the system administrator.

Workload Manager

The system administrator classifies work to service classes. The classification mechanism uses work attributes like transaction names, user identifications or program names which specific applications are known to use. In addition the system administrator defines goals and importance levels for the service classes representing the application work. The goals define performance expectations for the work. Goals can be expressed as response times, a relative speed (termed velocity) or as discretionary if no specific requirement exists. The response time describes the duration for the work requests after they entered the system and until the application signals to WLM that the execution is completed. WLM is now interested to assure that the average response time of a set of work requests ends in the expected time or that a percentage of work requests fulfill the expectations of the end user.

Workload Manager

The definition of a response time also requires that the applications communicate with WLM. If this is not possible a relative speed measure – named execution velocity - is used to describe the end user expectation to the system.

Workload Manager

This measurement is based on system states which are continuously collected. The system states describe when a work request uses a system resource and when it must wait for it because it is used by other work. The latter is named a delay state. The quotient of all using states to all productive states (using and delay states) multiplied by 100 is the execution velocity. This measurement does not require any communication of the application with the WLM component but it is also more abstract than a response time goal.

Finally the system administrator assigns an importance to each service class to tell WLM which service classes should get preferred access to system resources if the system load is too high to allow all work to execute. The service classes and goal definitions are organized in service policies together with other constructs for reporting and further controlling and saved as a service definition for access to WLM. The active service definition is saved on a couple data set which allows all z/OS systems of a Parallel Sysplex cluster to access and execute towards the same performance goals.

WLM is a closed control mechanism which continuously collects data about the work and system resources; compares the collected and aggregated measurements with the user definitions from the service definition and adjusts the access of the work to the system resources if the user expectations have not been achieved. This mechanism runs continuously in pre-defined time intervals. In order to compare the collected data with the goal definitions a performance index is calculated.

The performance index for a service class is a single number which tells whether the goal definition could be met, has been overachieved or was missed. WLM modifies the access of the service classes based on the achieved performance index and importance. For this it uses the collected data to project the possibility and result of a change. The change is executed if the forecast comes to the result that it is beneficial for the work based on the defined customer expectations. WLM uses a data base ranging from 20 seconds to 20 minutes to contain a statistically relevant basis of samples for its calculations. Also in one decision interval a change is performed for the benefit of one service class to maintain a controlled and predictable system.