Simulation methodology and tools for the development of novel program execution models and architectures
Author(s) | Pavel, Robert | |
Date Accessioned | 2015-10-22T12:38:26Z | |
Date Available | 2015-10-22T12:38:26Z | |
Publication Date | 2015 | |
Abstract | The exascale era of high performance computing will be defined by novel architectures with a focus on power efficiency and resiliency as well as program execution models with a focus on usability and scalability. To develop and evaluate these architectures and program execution models, simulation and modelling tools must be used. However, existing simulation and modelling tools are very complex and tend to emphasize high performance simulations. And to achieve these speeds, two general methods are used. Many simulators simply eschew accuracy in the name of performance. Others still are the product of large amounts of time and effort to optimize the simulation of a specific class of architecture or system, which is perfect for studying existing systems and evolutionary designs but not for studying revolutionary designs. To this end, I have developed PICASim as a model and framework for the development and study of novel architectures and program execution models. The PICASim model's compromise between accuracy and performance can be adjusted during the early development of exascale systems to obtain highly accurate simulations of smaller, but representative systems to aid in the study of the impact that new architectural features and runtime capabilities will have on the performance of a full exascale system. Additionally, its modular nature and focus on composability will allow for new architectural features to be evaluated and studied with ease. Additionally, I have developed a language to express system graphs and the interconnection of components at a high level: the LADS. The LADS is a language used to express PICASim graphs but is already being extended to support codelet-based languages as well. The LADSpiler is a source-to-source compiler designed to transform a graph expressed in the LADS to source code with all interconnections represented and additional optimizations, based on graph theory, applied. The contributions of my dissertation will be the following: 1. PICASim: A model for Parallel Discrete Event Simulation that combines the flexibility of the Chandy and Misra model with the modularity and intuitiveness of the Bryant model 2. The PICASim Framework: A Fully-Distributed PDES framework designed to study and develop novel architectures and program execution models at an indicative scale prior to devoting resources toward full-scale simulation 3. Results demonstrating PICASim's ability to model said architectures and program execution models with a high degree of accuracy and flexibility. 4. Results demonstrating PICASim's use in the study and development of novel program execution models and architectures 5. The LADS: A language used to express system graphs, including but not limited to PICASim graphs 6. The LADSpiler: A source-to-source compiler that can optimize and translate systems expressed in LADS to the target language. | en_US |
Advisor | Gao, Guang R. | |
Degree | D.Eng. | |
Department | University of Delaware, Department of Electrical and Computer Engineering | |
DOI | https://doi.org/10.58088/e7qv-2c91 | |
Unique Identifier | 926040372 | |
URL | http://udspace.udel.edu/handle/19716/17186 | |
Publisher | University of Delaware | en_US |
URI | http://search.proquest.com/docview/1708653992?accountid=10457 | |
dc.subject.lcsh | Computer programs -- Execution -- Computer simulation. | |
dc.subject.lcsh | Computer architecture -- Computer simulation. | |
dc.subject.lcsh | Graph theory. | |
Title | Simulation methodology and tools for the development of novel program execution models and architectures | en_US |
Type | Thesis | en_US |