Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Date of Award

9-2011

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Electrical and Computer Engineering

First Advisor

Sandip Kundu

Second Advisor

Maciej Ciesielski

Third Advisor

Israel Koren

Subject Categories

Computer Engineering | Computer Sciences

Abstract

Manufacturing technology has permitted an exponential growth in transistor count and density. However, making efficient use of the available transistors in the design has become exceedingly difficult. Standard design flow involves synthesis, verification, placement and routing followed by final tape out of the design. Due to the presence of various undesirable effects like capacitive crosstalk, supply noise, high temperatures, etc., verification/validation of the design has become a challenging problem. Therefore, having a good design convergence may not be possible within the target time, due to a need for a large number of design iterations.

Capacitive crosstalk is one of the major causes of design convergence problems in deep sub-micron era. With scaling, the number of crosstalk violations has been increasing because of reduced inter-wire distances. Consequently only the most severe crosstalk faults are fixed pre-silicon while the rest are tested post-silicon. Testing for capacitive crosstalk involves generation of input patterns which can be applied post-silicon to the integrated circuit and comparison of the output response. These patterns are generated at the gate/ Register Transfer Level (RTL) of abstraction using Automatic Test Pattern Generation (ATPG) tools. In this dissertation, anInteger Linear Programming (ILP) based ATPG technique for maximizing crosstalk induced delay increase at the victim net, for multiple aggressor crosstalk faults, is presented. Moreover, various solutions for pattern generation considering both zero as well as unit delay models is also proposed.

With voltage scaling, power supply switching noise has become one of the leading causes of signal integrity related failures in deep sub-micron designs. Hence, during power supply network design and analysis of power supply switching noise, computation of peak supply current is an essential step. Traditional peak current estimation approaches involve addition of peak current associated with all the CMOS gates which are switching in a combinational circuit. Consequently, this approach does not take the Boolean and temporal relationships of the circuit into account. This work presents an ILP based technique for generation of an input pattern pair which maximizes switching supply currents for a combinational circuit in the presence of integer gate delays. The input pattern pair generated using the above approach can be applied post-silicon for power droop testing.

With high level of integration, Multi-Processor Systems on Chip (MPSoC) feature multiple processor cores and accelerators on the same die, so as to exploit the instruction level parallelism in the application. For hardware-software co-design, application programming model is based on a Task Graph, which represents task dependencies and execution/transfer times for various threads and processes within an application. Mapping an application to an MPSoC traditionally involves representing it in the form of a task graph and employing static scheduling in order to minimize the schedule length. Dynamic system behavior is not taken into consideration during static scheduling, while dynamic scheduling requires the knowledge of task graph at runtime. A run-time task graph extraction heuristic to facilitate dynamic scheduling is also presented here. A novel game theory based approach uses this extracted task graph to perform run-time scheduling in order to minimize total schedule length.

With increase in transistor density, power density has gone up substantially. This has lead to generation of regions with very high temperature called Hotspots. Hotspots lead to reliability and performance issues and affect design convergence. In current generation Integrated Circuits (ICs) temperature is controlled by reducing power dissipation using Dynamic Thermal Management (DTM) techniques like frequency and/or voltage scaling. These techniques are reactive in nature and have detrimental effects on performance. Here, a look-ahead based task migration technique is proposed, in order to utilize the multitude of cores available in an MPSoC to eliminate thermal emergencies. Our technique is based on temperature prediction, leveraging upon a novel wavelet based thermal modeling approach.

Hence, this work addresses several optimization problems that can be reduced to constrained max-satisfiability, involving integer as well as Boolean constraints in hardware and software domains. Moreover, it provides domain specific heuristic solutions for each of them.

Share

COinS