National Technical University of Athens (NTUA)
Performance variation in digital systems: workload dependent modeling and mitigation
Abstract
dc:descriptionThe computer industry is witnessing an unprecedented demand for more functionality and performance and is continuously using silicon components with smaller form factor and feature size. This aggressive downscaling of hardware components has brought about several failure mechanisms that degrade the system’s operation, threatening its dependability. Such failure mechanisms can be the result of the natural occurring variation in the attributes of circuit elements during the fabrication procedure, or can be attributed to the aging and the gradual wearout of the hardware and other variability effects related to space particles and power/ground line voltage variation. The inherent stochastic nature of these failure phenomena contributes to the so-called performance variation of digital systems, in the sense that system behavior and response cannot be fully deterministic and have a dynamic component.In the software layer, computational- and data-intensive applications, user interaction and quality of service conditions also generate persistently varying and unpredictable workloads, deteriorating this effect. While software applications are becoming even more complex and resource-hungry (especially due to the continued “virtualization” that leads to the ubiquitous use of run-time threads and dynamic memory allocation) and since the shrinking of transistor and interconnect dimensions is not expected to end in this decade, we can assume that we have already entered an era of inevitable, strongly dynamic performance variation.Under this highly dynamic context of system operation, ensuring dependability and meeting timing constraints seems challenging. The goal of the current research is to study existing methodologies that mitigate performance variation and develop related schemes that can ultimately ensure and guarantee timing deadlines. For this reason, first the dominant failure mechanisms which create defects is the silicon layer and deteriorate the reliability of the system are presented. A thorough review of the prior art on the subject of reliability mitigation is also shown, in order to realize the current, state-of-the-art mitigation approaches and methodologies. Then, the aforementioned reliability threats need to be captured and modeled while their impact on the system’s performance should be described and estimated. Hence, existing tools are employed, and new ones are suggested, in order to develop a complete framework that effectively evaluates the failure probability of electronic components, focusing especially on the SRAM buffers of NoC routers. Later, a realistic case study of a closed-loop PID controller is presented to mitigate performance variation with a reactive DVFS response. This scheme is compared against the version of a Linux CPU frequency governor in terms of energy consumption and timing response. Moreover, another flavor of this scheme is presented, to perform thermal management. Again this controller is implemented on pure hardware and illustrated with a realistic case study.Next, the aforementioned PID controller is improved to operate on finer granularity, at the thread node level. The concepts of performance and deadline vulnerability factor are introduced to support the formulation of a discrete time control problem while the basis of this new approach utilizes the system scenario methodology; this methodology, along with related terms and definitions, is studied in detail. In addition, a run-time adjustment on this methodology to adapt to performance variability norms is shown, creating an adaptive scenarios scheme and achieving notable energy gains. Still, however, performance variation is managed with a reactive response and no timing guarantees are yet delivered.Finally, proactive DVFS actuations on the thread node level are proposed, using dynamic scenarios to guarantee timelines in a cost-efficient manner. By exploiting the partial predictability of the application behavior, a dynamic scenario approach is developed to enable cost-effective DVFS decisions. Simulation results present significant energy gains compared to previous frequency guardband methods while experimental results on the hardware platform substantiate the effectiveness of our scheme.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Noltsis, Michail
- Νόλτσης, Μιχαήλ
Subjects
dc:subject × 8- MOS αξιοπιστία διατάξεων
- Circuit reliability
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/49537