Formulating a statistical model of tool runtimes

[ssi...@gmail.com]

Hi,
I work as a graduate student in computer science, and I am
interested in developing a model to estimate runtimes for a particular
tool I am using. For each run of the tool, I have available to me
several attributes about the program ( lines of code, number of loops,
etc) Let's say I have information about N attributes. I have run a
number of programs through my tool and have generate tool runtimes for
those programs. What I want to be able to do (if possible) is to
estimate the runtime of the program using only the information I have
a priori ( lines of code, number of loops, etc). I have a feeling
there are some statistical techniques that will enable to capture of
model of the runtime of the program based on these known attributes,
and was hoping someone might be able to point me in the right
direction for creating such a model. Thanks for your help,

Scott

[Gus Gassmann]

Your first step should probably be some exploratory data analysis,
starting with some plots. How do you see the runtimes increase with
the number of lines of code, for instance? Is there an apparent linear
or nonlinear relationship? Same for other relationships. Then you
should formulate a theoretical model, something like

runtime = f(attribute_1, attribute_2,...,attribute_N) + error

In all likelihood your first model will be linear

runtime = a_0 + a_1*attribute_1 +a_2*attribute_2 +...+ a_N*attribute_N
+ error

You estimate the coefficients using a regression package and... Bob's
your uncle. Details about this ought to be in pretty much any
introductory stats book.

Cheers