Optimization and Regulation of Performance for Computing Systems

The current demands of computing applications, the advent of technological advances related to hardware and software, the contractual relationship between users and cloud service providers and current ecological demands, require the re\ufb01nement of performance regulation on computing systems. Powerful mathematical tools such as control systems theory, discrete event systems (DES) and randomized algorithms (RAs) have o\ufb00ered improvements in e\ufb03ciency and performance in computer scenarios where the traditional approach has been the application of well founded common sense and heuristics. The comprehensive concept of computing systems is equally related to a microprocessor unit, a set of microprocessor units in a server, a set of servers interconnected in a data center or even a network of data centers forming a cloud of virtual resources. In this dissertation, we explore theoretical approaches in order to optimize and regulate performance measures in di\ufb00erent computing systems. In several cases, such as cloud services, this optimization would allow the fair negotiation of service level agreements (SLAs) between a user and a cloud service provider, that may be objectively measured for the bene\ufb01t of both negotiators. Although DES are known to be suitable for modeling computing systems, we still \ufb01nd that traditional control theory approaches, such as passivity analysis, may o\ufb00er solutions that are worth being explored. Moreover, as the size of the problem increases, so does its complexity. RAs o\ufb00er good alternatives to make decisions on the design of the solutions of such complex problems based on given values of con\ufb01dence and accuracy. In this dissertation, we propose the development of: a) a methodology to optimize performance on a many-core processor system, b) a methodology to optimize and regulate performance on a multitier server, c) some corrections to a previously proposed passivity analysis of a market-oriented cloud model, and d) a decentralized methodology to optimize cloud performance. In all the aforementioned systems, we are interested in developing optimization methods strongly supported on DES theory, speci\ufb01cally In\ufb01nitesimal Perturbation Analysis (IPA) and RAs based on sample complexity to guarantee that these computing systems will satisfy the required optimal performance on the average.

Name thesis:degree_name

Electrical Engineering

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical and Computer Engineering

Year

2015

Author dc:creator

• Luna Castaneda, Jose Marcio

Contributors dc:contributor

• Abdallah, Chaouki
• Heileman, Gregory
• Fierro, Rafael
• Lorenz, Jens

• Computing Systems
• Randomized Algorithms
• Discrete Event Systems
• Infinitesimal Perturbation Analysis
• Statistical Learning

Language dc:language

English

Repository record dc:identifier

https://digitalrepository.unm.edu/ece_etds/163

OAI identifier oai:identifier

oai:digitalrepository.unm.edu:ece_etds-1162