University of Thessaly (UTH)
Scheduling and performance characterization on heterogeneous computing systems
Abstract
dc:descriptionThe era of performance and power efficiency scaling through transistor shrinking and frequency scaling, characterized by Dennard’s law, has come to an end. Nevertheless, transistor doubling still withstanding, computing architectures have moved to the multicore and manycore era for improving performance through parallelism. However, this scaling of computing architectures exposed power consumption to be an important limiting factor and a key concern to sustain further growth. Heterogeneous computing architectures are the next evolution step and promise to improve the performance, power and cost efficiency of systems. They promise to do so through hardware specialization, by including components which exhibit different performance and power consumption characteristics. The potential for improved performance and power efficiency from heterogeneous architectures has rendered them ubiquitous in every aspect of computing, from mobile devices such as smartphones and tablets to large-scale computing installations such as supercomputers and datacenters. However, the efficient allocation of heterogeneous resources to match workload requirements with their capabilities, thus achieve their performance and power efficiency potential, is an open and challenging problem.In this dissertation we contribute novel solutions, techniques and insight on efficient resource allocation for heterogeneous systems. At the level of heterogeneous processor architectures, we present new techniques for managing instruction-based heterogeneity on state-of-the-art, shared-ISA architectures. These include binary adaptation techniques to enable software portability and thread migration across heterogeneous cores despite their binary incompatibilities. Following, we present a novel, heterogeneous-aware scheduler for maximizing the speedup of a workload through dynamic re-allocation of heterogeneous cores between co-executing programs. The evaluation of our methods on a hardware prototype platform shows that the dynamic scheduler we propose outperforms significantly previous state-of-the-art solutions. Moreover, we scope resource allocation for performance and energy efficiency on heterogeneous datacenter architectures. We present new methodologies and define iso-comparison metrics to fairly and accurately characterize the performance and energy efficiency across diverse, heterogeneous server platforms in the context of online data analytic services. We evaluate several heterogeneous types of servers, including standard servers, micro-servers and manycore accelerators. Our methods provide new insight by assessing resource allocations in heterogeneous datacenter architectures under the prism of meeting Quality-of-Service requirements of online data analytics while characterizing their energy efficiency. The application of our methodologies can effectively guide efficient heterogeneous resource allocation and provisioning in the datacenter.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Georgakoudis, Giorgis
- Γεωργακούδης, Γιώργης
Subjects
dc:subject × 20- Eτερογενή συστήματα υπολογιστών
- Χαρακτηρισμός απόδοσης
- Χρονοπρογραμματισμός
- Δυναμική επανεγγραφή δυαδικού κώδικα
- Μικρο-εξυπηρετετητές
- Ενεργειακή αποδοτικότητα
- Heterogeneous computing systems
- Performance characterization
- Scheduling
- Dynamic binary rewriting
- Micro-servers
- Energy efficiency
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Computer and Information Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/40218