Examples of the Exception Analysis Process

The capacity planner of a large enterprise wants to determine whether the MIS Business Intelligence servers (misbi4a, misbi5a, and misbi6a) in her company are underused. These servers are dedicated to the MIS group. The capacity planner wants to use facts to convince the MIS group to allow these BI servers to be shared by other teams in the organization that require access to the Business Intelligence software that is hosted on these servers. The following example describes how this is accomplished.

This example uses a filter to specify MIS business intelligence servers. To avoid seeing the same exceptions every day, the exception specification uses a filter to look for exceptions for the last day only. The input source is a DAYHOUR table called DayHourSystem. The exception definition is set up to flag an exception if it discovers six consecutive hours where the CpuBusyPctWMean is less than 20%. The group that is specified in the Exception transformation for this example is MACHINE.

For this example, the Exception transformation (CPU Utilization For MIS Business Intelligence Servers) is executed from the job called System Daily Exceptions. Both the exception table and the exception condition table are specified for this transformation. Their contents can be used for additional processing.

The name of the Exception transformation is specified on the General tab. In this example, the name is CPU Utilization For ITRM Servers.

The Filters tab causes the transformation to process only observations that pertain to the date of the last date of data and the three machines (misbi4a, misbi5a, and misbi6a).

The filters that are specified on the Report Attributes tab are shown in the following display.

The Domain subcategory is not specified.

The Group by value for this example is MACHINE. The Order Column is DAYDATE.

As shown in the following display, the Exception Type is Constant threshold. The Number of observations in a row is set to 6.

When the job that contains the Exception transformation is run and if an exception is detected, two types of reports are generated:

For this example, both the exception table and the exception condition table are output from the transformation.

The Exception transformation creates an overview report. If exceptions are found, the transformation also generates individual reports. The following display shows the unexpanded overview report.

To see the expanded version of the overview report, click the name of the exception definition. (In the preceding display, this name is circled.) The expanded overview report lists the Group By values (for the MACHINEs) that matched the condition that was specified by the exception definition. In this example, the following MACHINES had CPU Utilization below 20% for 6 observations (hours) in a row: misbi4a, misbi5a, and misbi6a.

To see the individual reports for the MACHINES that experienced low CPU utilization, click the corresponding MACHINE. In the preceding display, the misbi4a MACHINE is circled. That action displays the individual report for misbi4a.

To see the observations that matched the expression that was specified in the exception definition, click Link to Observations That Match the Exception Definition Expression Report. (The link is circled in the preceding display.) The expression that was specified in the exception definition is CPUBUSYWMEAN < 20.

As shown in the individual report for misbi4a, the server is consistently underused. Except for two significant periods of moderate usage, the server is not experiencing much activity.

The recommendation that is circled in this display suggests that this server be considered for consolidation with another server.

The capacity planner and the performance analyst would both like to be warned of sudden spikes in resource consumption. Volatility of demand is a common occurrence to some extent. However, sudden rises in resource demand might also be due to changes in customer behavior or business needs that were not anticipated. It might be necessary to make adjustments on a short-term basis, and the capacity plan might need to be modified as well.

In this example, the performance analyst does not want to set a fixed threshold, because a slow and steady growth in demand is expected. However, the analyst still wants to be notified if the growth in demand is more sudden. To accomplish this for CPU consumption, the analyst first defines a moving average of the period and a moving standard deviation for CPU time in the aggregation summary table. (These are based on the weighted mean of CPUBUSY.) These statistics are used to construct a dynamic (rather than static) threshold for warning of spikes in demand.

In a new job with an Exception transformation, using this aggregation summary table as input, the analyst defines an exception definition that can detect sudden spikes in CPU consumption. The exception definition uses statistic bounds to compare the current CPU consumption as a weighted mean (CPUBUSYWMEAN) with a dynamic threshold consisting of the moving average of the same underlying column (CPUBUSYWMEAN_MA) plus or minus two standard deviations for the moving average (CPUBUSYWMEAN_MSD).

In the Exception transformation itself, the analyst specifies that reports be generated and that an alert be sent if an exception is found.

If an exception is detected when the job that contains the Exception transformation is run, two types of reports are generated:

For this example, both the exception table and the exception condition table are output from the transformation.

The following individual report is generated for this example:

The analyst tested the new Exception transformation and definition by executing it against existing data because she had already encountered recent episodes of spikes in demand.

An exception was generated for the misbi5a machine, among others, during the period for which the analyst already had data. As illustrated by the plot, CPU consumption exhibited slow but steady minor growth until the week of June 16. The statistic bounds formed by the moving average and moving standard deviation adjust for this growth automatically. However, during the week of June 16, CPU consumption exploded upward and exceeded the dynamic threshold. The exception was triggered by this sudden spike in demand.

Satisfied that this exception definition would provide her with alerts and accompanying reports for the next spike in demand, the analyst put this job into the normal daily schedule. The next time that this type of sudden increase in demand occurs, the analyst will be notified on a daily basis alerted of any future occurrence.