The RASSP Digest - Vol. 3, September 1996

EaSE Trades Technical Review

by Gary W. Panzer

Abstract

A methodology for the rapid, systematic, through and efficient exploration of very large digital processing design trade spaces has been developed. The resulting process is implemented as an Early System Evaluation and Trades (EaSE Trades) Design Advisor Tool. The Navy's Processing Graph Methodology (PGM) and JRS Integrated Design Automation System (IDAS) combined with the use of Design of Experiment (DOE) techniques for the experimental exploration of candidate design alternatives have achieved breakthrough improvements in the way trade studies are accomplished.

1. Objectives

The objective of RASSP is to improve the process by which embedded digital signal processors (DSP) are designed, manufactured, upgraded, and supported. This initiative has top-level goals of developing an advanced, systematic design capability to achieve

The RASSP design system relies heavily on virtual prototyping, that is, extensive simulation at increasing levels of fidelity. DSP system and architecture simulation tools, such as the JRS NETSYN are being developed to meet the virtual prototyping goals. These tools offer powerful methods for examining system alternatives early and often in the design process. They can be applied on problems covering a wide range of architectures, applications, and rates.

The capability of these new tools opens doors to exploration of vast alternate trade spaces. Even with the improvement in speed offered by these tools, thorough trade studies using simulation tools often cannot examine all potentially interesting trade conditions. Because of cost and time constraints, we are often forced to narrow the number of cases using engineering judgment or other subjective criteria. Important alternatives may be excluded before the trade analysis even starts. The solution lies in the use of a systematic methodology to search the trade space of alternatives.

In many commercial industries, the challenge of intelligently sampling a large trade space is accomplished using the field of statistics commonly referred to as Design of Experiments (DOE). By definition, DOE is the planned, structured, and organized observation of input (independent) variables (hereafter referred to as factors), and their effect on the output (dependent) variables (hereafter referred to as responses). In many industries and applications, DOE analysis has resulted in a considerable reduction in data collection necessary for exploration of large candidate trade spaces.

The Early System Evaluation and Trades (EaSE Trades) Program is developing a methodology for the rapid, systematic, thorough, and efficient exploration of large digital signal processing trade spaces. The process uses DOE as the basis for a Design Advisor that aids the engineer in the selection of simulation cases that will help to find the best trade choice quickly and efficiently.

2. Technical Approach

The EaSE Trades method is shown in Figure 1. It outlines the step-by-step process for rapidly searching the large, multidimensional design trade space of the DSP system. Sampling is done with confidence that all the potentially applicable trade conditions will be considered. The following paragraphs detail this process.

Screening Matrix Selection - In most physical situations, some variables will be much more important and effective than others at meeting the customer's needs. DOE screening studies take advantage of this to quickly, and systematically, reduce the number of simulation cases that need to be examined. The user specifies the factors to be traded, the responses of interest, and any constraints on the problem. The EaSE Trades method uses sampling techniques designed for screening to intelligently choose simulation cases from the large trade space as represented by the factors being traded. The sampling plan recommends a matrix of test conditions to be simulated on NETSYN. The simulation results quantify the response values of interest.

Analysis and Visualization Refining Matrix Selection - The smaller trade space is then examined in a more refined experiment. DOE sampling techniques which are designed to produce a second order polynomial description of the response surface are used to create a matrix of test conditions to be simulated on NETSYN. A second order regression equation for each response of interest is derived from the refined experiment.

Analysis and Adequacy Check - The "goodness of fit" of the second order equation is tested statistically to assess the adequacy of the second order equation to describe the observed responses. If required, additional samples can be simulated to facilitate the use a higher order polynomial to describe the relations between factors and responses.

Design Optimization - Assuming a good fit, the second order polynomial equations are used to trade-off alternative solutions. The equations are solved for the solution that best meets the customer's needs. This optimum design is then simulated on NETSYN to confirm the predictions of the second order model.

The final product will be a confirmed optimum design for the DSP system.

3. Technical Results

Processing an experiment through the EaSE Trades Design Advisor produces results depicted in Table 1. This partial table lists the coefficients on an equation that predicts system performance. Coefficients of lesser importance have been removed from the table.

The processing system's response equation is in the form:

Y = B0 + B1X1 + B11X12 + B2X2 + B22X22 + B3X3 + B33X32 + B12X1X2 + B13X1X3 + B23X2X3 + B123X1X2X3

Examining the above table, the negative coefficient values for the linear terms indicate a decreasing effect on the response time when the number of processors is increased. This is in agreement with our intuitive expectation. Processors P11 and P31 also appear as the most important in the response time. This is also in agreement with our expectations since they are the fastest processors as indicated by their specifications. Further analysis determined that the data behaved inverse square linearly, i.e., the reciprocal of the square of the response time gives the best fit. This confirms our original understanding of the system's 'inverse exponential' behavior.

4. Deliverables

The major deliverables of the program are

Methodology Definition Document - This material covers the technical aspects of the system operation for this program. It defines the overall strategy for conducting the research, and the approach to controlling the experiments and evaluating the results.

Integration and Test Plan - This documents the specific plan, including schedule, for mechanizing the tool interfaces, debugging and testing for proper operation prior to conducting methodology evaluations in the validation phase of the program.

Methodology Evaluation Plan - This is a specific plan how evaluations are to be conducted. It is intended to provide for a cost-controlled and disciplined research approach.

Final Report - This is a program and technical report that summarizes the work accomplished. Other deliverables will be included by reference or as appendices as specified by the contract.

Design Advisor Prototype - The Design Advisor software and documentation that is in-place at the conclusion of the research will be made available.

Gary W. Panzer
Hughes Radar and Communication Systems
P.O. Box 92426, RE/R01/A528
Los Angeles, CA 90009-2426
panzer@bala.hac.com

Newsletter Index
The RASSP Digest - Vol. 3, September 1996
newsletter/html/96sep/news_5.html