National Technical University of Athens (NTUA)
Τεχνικές αποδοτικής εικονικοποίησης επιταχυντών για νεφοϋπολογιστικά περιβάλλοντα
Abstract
dc:descriptionAs data creation worldwide in today's world keeps growing with remarkable rates, the processing power needs to be increased proportionally. However, the multicore scaling trend presents a limit in the foreseeable future, which is referred in the literature as the dark silicon era. In this context, computer scientists and professionals estimate that the way data are processed will adapt to the new conditions and future data centers will gradually move from the scale-up paradigm to more heterogeneous architectures embracing accelerating resources. At the same time, cloud computing has been established in many data center infrastructures offering benefits both for the end users as well as the service providers, such as flexibility, server consolidation, cost reduction and better resource utilization among others. With the aforementioned potential that heterogeneous computing and accelerators appears to develop, there is a growing need for integration in the current cloud stacks. In essence, virtualization-aware systems need to embrace accelerators by adapting their components into the specialized nature of this kind of hardware. In this thesis, we explore the implications of integrating accelerator devices into the virtualization ecosystem. We identify the key challenges of virtualizing accelerator resources and we outline the reasons that the traditional I/O virtualization methods are not adequate for this kind of specialized devices. We explore virtualization techniques targeting two popular accelerator families, NVIDIA GPUs and Intel Xeon Phi coprocessors, following different approaches based on the nature of each accelerator environment. In our GPU virtualization approach, we propose the use of a remote acceleration framework in a single-node virtualization platform combined with a low overhead intra-node framework which results in efficient application offloading in virtualized environments. The data path in the design of our intra-node framework is realized through the hypervisor as the network medium, instead of the driver domain. Furthermore, we evaluate our prototype using both network microbenchmarks and analyzing a common GPU stencil, showing that it is a viable approach for accelerator software stacks that are released in a closed manner. Specifically, evaluation results show that our approach boosts the transfer throughput by a factor of up to 6.3 compared to the remote default path, while it adds an overhead of 15% in terms of GPU execution compared to direct device assignment configuration. Regarding our Xeon Phi virtualization approach, to the best of our knowledge it is the first and currently the only solution that enables sharing of a Xeon Phi device by multiple virtual machines running on the same physical node. We follow the paravirtualization technique targeting the low-level transport layer of the software stack. In this way, the design principles of our approach can be applied to future accelerator technologies as well. Evaluation shows that our prototype can enable better accelerator utilization when it is used by multiple VMs, increasing up to 3.56x the total throughput versus a single host application, which represents the direct device assignment configuration.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Gerangelos, Stefanos
- Γεράγγελος, Στέφανος
Subjects
dc:subject × 12- Εικονικοποίηση επιταχυντών
- Διαμοιρασμός επιταχυντών
- Εικονικές μηχανές
- Νεφοϋπολογιστικά περιβάλλοντα
- Accelerator virtualization
- Accelerator sharing
- Virtual machines
- Cloud environments
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/46344