HP 3000 Manuals

HP RXForecast Users Manual for MPE Systems

Examining a Case Using Business Units 

Adding a system to be served by Trapper or reorganizing the structure of
the mail network could have serious effects on the performance of Trapper
as a mail hub.  The problem could cause serious delays in mail leaving
and entering the hub.

Performance is acceptable at present.  Batch jobs that run at night
always finish within the designated time.  There are no guarantees that
this will continue, however, since the mail system is changing.

The plan is to upgrade Trapper from a Series 58 to a Series 70 in
February 1989.  Meanwhile, users will slowly be added to the systems that
use Trapper as a mail hub.

The objective of the forecast is to determine if performance will remain
at an acceptable level until the upgrade.  You need to know this before 
it becomes too late to revise the schedule or restructure the mail
system.

RXForecast Options Selections 

For this example, set options in the RXForecast Options command dialog
box as follows:


            

   *   Seasonality Level.

       Check all Seasonality Levels.

   *   Confidence Level.

       Set the confidence level to 90% Confidence.

   *   Interval Type.

       Choose Confidence Interval.

Forecast Selections 

As Global CPU Utilization approaches 90 percent, batch jobs will probably
not finish during their scheduled time, and they will start to compete
with other jobs and sessions during the day.  This situation could easily
lead to delays in the mail system.

   *   Metric.

       If you can track Global CPU Utilization for the time that the mail
       system is most active, you could determine if significant mail
       delays would occur.

   *   Dates.

       All of the logged data is considered valid.  No major changes have
       occurred during the past 6 months, so you will use all of the
       data.  Forecast the next 6 months and note the projected CPU
       Utilization.

   *   Shift.

       Choose a shift from 9:00 PM to 7:00 AM.

   *   Ignore Weekends.

       Ignore weekends since the most critical periods occur during the
       week.

   *   X-Axis.

       Six months of data along with a 6-month forecast means choosing an
       X-Axis of Year.

   *   Points Every.

       Choose Points Every Day.

   *   Trending Method.

       Select the Linear method.  This may be the right choice, since you
       expect steady growth up until February.

   *   Show Threshold.

       Change the Show Threshold option on the Forecasts command NEXT
       dialog box to place a threshold line on the graph at 90 percent
       because you are concerned about when CPU Utilization will exceed
       90 percent.

Resulting Graph 

Generate the forecast.  The following graph is the result:


            

Analysis 

The forecast is difficult to read.  It does not appear valid, however,
because there seems to be a lot of variability on the graph.  Also, the
line plotted on the graph that corresponds to the forecast does not
appear to be the best trend since all of the data points are widely
dispersed around the trend line.  Finally, when you look at the Stats
report, you see that R-Squared 0.31 indicates a poor correlation.

Reforecast 

You can draw a forecast that is easier to read.  Return to the first
steps in the forecasting process.

Forecast Selections 

   *   Points Every.

       Change the summarization from Points Every Day to Points Every
       Month.  This change could eliminate some of the dispersion, but
       still give you sufficient detail to make a decision.

Resulting Graph 

Run the model and generate the following graph:


            

Analysis 

Immediately you notice that the trend does not fit the actual data.  To
confirm this impression, generate the Stats report.

     Regression Analysis for Trapper. Metric = CPU Util.
     DATE : 02/26/90   TIME : 11:44:00 METHOD is Linear.
     Seasonality:  DAY  WEEK  MONTH  QUARTER  Auto Season ON.
     Time Window Start : 21:00   Stop : 07:00
     Date Window Start : 03/01/88   Stop : 08/25/88  Validate : 08/25/88
     Summarization: MONTH  Ignore Weekends:  ON.

     Model : FC = INT + SLOPE * TIME + DOW + WOM + MOQ + QOY
       FC=Forecast  INT=Intercept  TIME=# of time periods since start of forecast.
       DOW=Day of Week Seasonality  WOM=Week of Month Seasonality
       MOQ=Month of Quarter Seasonality  QOY=Quarter of Year Seasonality

     Trend line parameters:
         Intercept = 5.9585e+001;      T-Stat = 7.9      T-Prob. = 99.9
         Slope     = 3.9514e-001;      T-Stat = 0.2      T-Prob. = 15.2

     MSE = 65.30; Std. Err. = 8.08; R-Squared = 0.01; N = 6; P = 2;

     Day of Week Seasonality Parameter Estimates:
         Excluded: Not Statistically Significant

     Week of Month Seasonality Parameter Estimates:
         Excluded: Not Statistically Significant

     Month of Quarter Seasonality Parameter Estimates:
         Excluded: Not Statistically Significant

     Quarter of Year Seasonality Parameter Estimates:  (Calendar Quarter)
         Excluded: Not Statistically Significant

From the Stats report, you notice the following:

   *   The R-Squared statistic is almost 0.

   *   The Standard Error is fairly high at 8.08.

   *   The T-Probability for the slope is low at 15.2.

All measures indicate that the Linear method is not correct for this
situation.  If you tried either of the other two trend lines--S-Curve and
Exponential--you would find that they fit no better.

Business Unit File 

You have not yet investigated using the Business Units trending method.
You know that the system is heavily influenced by the HP DeskManager
application.  If you can somehow find a metric that captures the amount
of work done by HP DeskManager, you could possibly use this method.

Assume that you calculate a business unit value that reflects messages in
and messages out for each month.  You project how many users would be
added during the coming months, and, based upon this, project the
business unit value.  You obtain this number for all months between March
1988 and February 1989 and store this information in a file called
TRAPPER.BUS.

This is the business unit file, TRAPPER.BUS:

     Month  Date  Year  Business Unit
     =====  ====  ====  =============
      03     01   1988      130
      04     01   1988      125
      05     01   1988      100
      06     01   1988      115
      07     01   1988      124
      08     01   1988      130
      09     01   1988      140
      10     01   1988      145
      11     01   1988      150
      12     01   1988      165
      01     01   1989      170
      02     01   1989      175

The second, or Date, field is constructed to reflect monthly
summarization.  See "Creating a Business Unit File"  for a detailed
explanation of the business unit file.


Tip  TRAPPER.BUS is not an ideal name for this business unit file.  A
     name that indicates its summarization level would be better.  If you
     had several business unit files to choose from on the NEXT screen,
     you might have difficulty remembering which one had the
     summarization level you wanted.


RXForecast Options Selections 

Since you are going to use the Business Units trending method, uncheck
all Seasonality Levels that are equal to or higher than the seasonality
level built into the business unit file.


            

Forecast Selections 

Select TRAPPER.BUS on the NEXT screen (Business Unit File Assignment
option) by highlighting its name and clicking on  OK.


            

When you select TRAPPER.BUS, its name appears on the bottom of the first
screen of the Forecasts command dialog box.  Its selection also means
that the Business Units trending method is enabled.

   *   Trending Method.

       Select the Business Units trending method.

   *   Start Date.

       Verify that the Start date is roughly the same as the first date
       in TRAPPER.BUS. Both are 1 March 1988.

   *   Points Every.

       Points Every should be Month since TRAPPER.BUS has monthly
       summarization.

Resulting Graph 

Generate the following forecast:


            

Validate 

This forecast produces much better results than the previous one.  From
the Stats report, you can see the following:

   *   The R-Squared statistic is now 90 percent.

   *   The Standard Error is down to 2.61.

   *   The T-Probability is up to 99.6.

These measures indicate that the model fits the data quite well.

Some Conclusions 

The forecast for December is nearly 90 percent, which is dangerously
close to the level of CPU Utilization that you expect would cause poor
performance of HP DeskManager.

Examine the upper confidence limit for that forecast and notice that in
December it is well above the 90 percent level for CPU Utilization.  If
allowed to continue without any changes, the situation would probably
deteriorate such that you would have difficulties in completing HP
DeskManager processing at night and would find it interfering with other
parts of Trapper's workload.

Based upon this projection, you reschedule the upgrade for October 1988.

MPE/iX 5.0 Documentation