next up previous contents
Next: Decomposition in Analytic Modeling Up: Analytical Modeling Previous: Statistical Models

Summary

There are two trade-offs associated with analytical modeling: (1) the amount of detail about the system, and (2) the difficulty of determining the parameter values.

The first trade-off is the amount of detail about the system, or level of abstraction, that is embedded in the model. The parameters themselves and their representation determine the level of abstraction. More parameters decreases the level of abstraction (i.e., more details in the model) and increases the modeling cost. Regarding representation of system parameters, scalar parameters are simpler than functions and statistical tools, but less flexible. Functions require that the programmer determines not only their coefficients, but also their shape. Statistical tools require a specialized knowledge that average parallel programmers do not have.

The second trade-off is the difficulty of determining the parameter values, function coefficients, or tool specifications (e.g., variable distribution parameters). If the parameters are too specific, they can be hard to capture due to a lack of suitable tools or knowledge about the application. If the parameters are too high-level, it is necessary to determine representative behaviors of the system.

Analytical models usually have low computational costs, except for some statistical tools that require simulation efforts. All of these techniques require that the user understand the parallel system and then express that understanding as model parameters, although the model's methodology may orient the user and facilitate the understanding task.



Wagner Meira
Wed Feb 28 15:37:47 EST 1996