9. MULTIVARIATE REGRESSION FORECASTING › 9.6 Case Studies › 9.6.1 Capture Ratio Case Study › 9.6.1.1 Control Parameters

9.6.1.1 Control Parameters

Using a short set of menus that are described in Section 1.5
of this guide, you will arrive at the Multivariate Regression
Forecasting screen shown in Figure 9-9.

Figure 9-9 shows the control parameters that we used to
generate the final multivariate regression model in this case
study.

/--------------------  Multivariate Regression Forecasting  --------------------\
|Command ===>                                                                   |
|Enter a ? in any data entry field for more information on valid values.        |
|Composing CA MICS Inquiry:  CAPRAT - Capture ratios on test workload           |
|                                                                               |
|Report title ===> Capture Ratios on test workload                              |
|                                                                               |
|Selection criteria:                                                            |
|             Dependent file  ===> TOT -                                        |
|           Dependent element ===> TOTCPUTM                                     |
|                  Start date ===> _________ (ddmonyyyy)                        |
|                       SYSID ===> SYS1                                         |
|                        Zone ===> 1_______                                     |
|                   No. weeks ===> ____      (1-9999)                           |
|       Specify CAPAPU Values ===> N         (Y/N)                              |
|Save forecast                ===> ___       (YES/NO/AGE)                       |
|Independent file             ===> WKL - WORKLOAD USAGE FILE                    |
|Independent elements         ===> BATCPUTM CICCPUTM IMSCPUTM TSOCPUTM ________ |
|                                  ________ ________ ________ ________ ________ |
|                                                                               |
|Confidence limits (percent)  ===> __        (70/90/95)                         |
|Min R-square improvement     ===> 0.001     (0.000-0.100)                      |
|Delete observations          ===> N         (Y/N/R)                            |
|                                                                               |
\------------------------------------------------------------------------------/

 Figure 9-9.  Capture Ratio Case Study Control Parameters

Note that in this case study the analyst previously generated
two resource element files.  The first, TOT, contains a data
element, TOTCPUTM, which represents the total systemwide CPU
time.  This file and this data element are specified on the
control screen as the Dependent file and the Dependent
element.  The second file, WKL, contains the data elements
that represent the CPU time consumed by each of the workloads
running on the system.  This file and the corresponding data
elements are specified as the Independent file and
Independent elements for this case study.  Note also that the
analyst previously generated and saved forecasts for each of
the independent data elements, using one of the forecasting
routines of this component.  These forecasts are required for
Multivariate Regression Forecasting to produce forecasts of
the dependent variable.

The analyst also specified that the study would focus on
SYSID SYS1 and Zone 1.  Because no value is specified for the
Confidence limits parameter, we will use the default value of
95%.

In this case study, a Minimum R-squared improvement value of
.001 is used for the final model.  Sometimes in the
evaluation of multilinear models, you may feel that the
default value of .05 does not quite provide a fine enough
distinction between models to produce the best multilinear
regression model.  Normally, a value of .01 should be
sufficient to make this distinction.

In this particular case study, the first evaluation of the
model using the default Minimum R-squared improvement value
resulted in values for m of Equation 8 in Section 9.6.1 that
appeared to be somewhat high (for example, greater than 2).
The analyst therefore attempted the same regression model,
but with a Minimum R-squared improvement value of 0.001. This
time the stepwise regression process produced a different
model that provided m values which were closer to the true
values.  The regression process may provide different models
with the same list of potential dependent variables, and the
same input data, but with different Minimum R-squared
improvement values.  This is due to there being a model of
marginally higher r-squared value that fits the data but
which the regression process discarded as being irrelevant
due to the default Minimum R-squared improvement value, but
which it will not discard with a lower value.

There can be a significant amount of subjective judgment
involved in the evaluation of regression models.  It is
highly unlikely that two analysts working independently with
the same set of data would produce the same capture ratio
coefficients.  This does not mean that either of them is
necessarily incorrect, since this method of capture ratio
determination claims only to be an estimation technique.
There is no generally accepted technique for determining what
the exact capture ratios are on a specific processor.  Since
the uncaptured processor time is by definition not
distributed to the various workloads by the operating system,
there can be no method of independently checking the results.
Thus, the results of any capture ratio analysis are only an
estimate.