Reporting CPU utilization

Find out how the total CPU that is consumed across virtual processors is reported.

Prior to V5R3, processor utilization was calculated as a percentage of the available CPU time. Collection Services reported, in the performance database files, the time used on each processor along with elapsed interval time. Users of this data, such as the Performance Tools reports and displays, needed to add up the time used on each processor to get the total system CPU that was consumed. The available CPU time was calculated as the number of processors in the partition multiplied by the duration of the data collection interval. Finally, the CPU time was divided by the calculated available time to get the utilization percentages.

The problem with the previous methodology is that all users of the data assumed whole virtual processors and depended on no changes to the configured capacities. Logical partitions with partial processor capacities and the capability to do dynamic configuration no longer worked with this methodology. Temporary solutions for minimizing the impacts of these problems included scaling the utilization of the system processors to what would be reported for a whole number of processors and cycling Collection Services when the configuration changed. Because the individual job CPU time was not scaled, the additional time was accounted for by reporting it as being consumed by HVLPTASK. The HVLPTASK task did not actually use CPU, but CPU time was shown to be consumed by HVLPTASK for accounting purposes. The CPU time charged to HVLPTASK scaled the amount of work that was done by real jobs, which resulted in the system CPU percent utilization going from 0 to 100 in direct proportion to the amount of customer work that was performed.

In V5R3, Collection Services reports the total CPU that is consumed and the total CPU that is available to the partition within the interval. The concept of HVLPTASK and CPU scaling to whole virtual processors in shared processor environments does not exist. Collection Services no longer cycles the collection when the configuration changes.

Collection Services now reports the total processor time that is consumed by the partition along with the amount of processor time that was available to be consumed within the partition, regardless of the number of virtual processors that are configured, the partition units that are configured, or how they changed during the interval. To calculate utilization, users of this data divide the reported CPU consumed by the available capacity. This method of calculating CPU utilization eliminates the increasingly error-prone task of computing available CPU time. CPU utilization that is calculated with these new metrics is accurate regardless of how many processing units (whole or fractional) exist, when the processing units changed, or how often the units changed.

Several reasons account for this change in calculating CPU utilization. One reason is that with scaling utilization for jobs or groups of jobs appeared to be much smaller than would be anticipated. This concept is demonstrated in the example that follows. Another reason is that a configuration change could make CPU reporting not valid. Traditionally, the number of CPUs was based on the value that was configured at the beginning of a collection and an IPL was needed to change it. When dynamic configuration was introduced, Collection Services cycled the collection to handle the configuration changes, which assumed that changes would be infrequent. However, the more frequent the change, the more cycling occurs. If the changes are too frequent, collecting performance data is not possible. Lastly, even if the proper configuration data were reported and used for every interval, you would not know what happened between the time the interval started and until it completed. Utilization would still be calculated incorrectly in any interval where there was one or more configuration changes.

Example

Partition A has a capacity of 0.3 processor units and is defined to use one virtual processor. The collection interval time is 300 seconds. The system is using 45 seconds of CPU (15 seconds by interactive jobs and 30 seconds by batch jobs). In this example, the available CPU time is 90 seconds (.3 of 300 seconds). The total CPU utilization is 50%.

Prior to V5R3, when the numbers were scaled, system CPU usage is reported as 150 seconds. 150 seconds divided by 300 seconds of interval time results in 50% utilization. The interactive utilization is 15 seconds divided by 300 seconds, which is 5%. The batch utilization is 30 seconds divided by 300 seconds, which is 10%. The HVLPTASK is getting charged with 35% utilization (150 seconds minus 45 seconds), or 105 seconds divided by 300 seconds. These percentages give us a total of 50%.

Beginning in V5R3, the 45 seconds of usage is no longer scaled but is reported as is. The calculated CPU time that is derived from the reported consumed CPU time divided by the reported available capacity is 50% (45 seconds divided by 90 seconds). The interactive utilization percentage is 17% (15 seconds divided by 90 seconds). The batch utilization percentage is 33% (30 seconds divided by 90 seconds).