Back to search

University of Thessaly (UTH)

Exploiting intrinsic hardware guardbands and software heterogeneity to improve the energy efficiency of computing systems

Abstract

dc:description

A key challenge for both current- and next-generation computing infrastructure deployments is to increase the delivered computational performance within stringent power budget constraints due to power delivery, cooling and cost-related concerns. Techniques such as transistor shrinking, frequency scaling, and parallelism exploitation have contributed to shaping the power efficiency envelope of today's computing systems. However, all of the aforementioned practices are bound to hit the power wall of CMOS technology. Another approach towards power efficiency is hardware-level specialization. Heterogeneous computing combines different architectures, each appropriate for specific computational patterns, within the same system. A typical example is General Purpose programming on Graphics Processing Units (GPGPU), which exploits GPUs to efficiently execute certain classes of embarrassingly parallel computations. Heterogeneous computing, though, comes at the expense of significantly increased programmer effort.At the same time, chip manufacturers are introducing redundancy at various levels of CPU design to guarantee fault-free operation, even for worst case combinations of non-idealities in process variation and system operating conditions. This redundancy partly translates to CPUs operating at a higher voltage than what is strictly required, in the form of voltage margins. However, these worst case scenarios may appear only rarely or even not at all during the lifecycle of a given machine. Consequently, the operating voltage setting is overly pessimistic for typical operating conditions and leads to excessive power dissipation, which hinders the effort towards improving the power efficiency of computing infrastructures. On the software side, programmers very often write code that solves problems at a significantly higher accuracy than actually required to meet the Quality of Service (QoS) requirements of the application. In the same vein, all parts of the code are treated as equally important, despite the fact that their contribution to the quality of the end result may vary significantly. As a result, developers tend to write programs that “over-compute”, a practice which results in lower energy efficiency. The quest for more energy-efficient systems may necessitate a transition from CMOS towards a more efficient semiconductor technology. However, until a new, commercially viable candidate technology appears, optimizing the energy efficiency of systems based on CMOS technology is a worthy undertaking. In this dissertation we focus on the exploitation of the intrinsic hardware guardbands and software heterogeneity. We design and develop mechanisms that reduce the CPU operating voltage and, thus, minimize the CPU power footprint by considering and exploiting the computational characteristics of the executed workload. Besides improving energy/power efficiency, these mechanisms can also be used to mitigate performance penalties induced on a power-constrained platform. We also evaluate the trade-off between quality of results and energy efficiency when combining heterogeneous computing with various approximation techniques. More specifically:Power capping is commonly used to regulate the limited power resources of computing infrastructure deployments. We investigate the impact of reducing CPU voltage margins on the efficiency of software- and hardware-based power capping mechanisms on a variety of commercial, off-the-shelf platforms. We observe that operation at reduced voltage margins can significantly improve the efficiency of existing power capping mechanisms in terms of performance, CPU and system node power, as well as CPU temperature. Based on our investigation, we introduce a CPU power capping mechanism that preserves performance in power-constrained environments by operating the CPU at reduced voltage margins to reduce frequency throttling. We show that CPU power capping at reduced voltage margins results in performance improvement by up to 64% and 24% on average, compared with Intel’s RAPL and Dynamic Frequency Scaling (DFS) mechanisms, respectively. Given that CPU operation at reduced voltage margins may potentially limit the effectiveness of a reliability safeguard introduced by manufacturers, we validate the robustness of our approach with a set of long-running experiments. We also show that significant energy gains can be achieved even when taking into account the cost of checkpointing and recovery in large-scale systems. Furthermore, we investigate the association of workload characteristics with the extent of exploitable reduction of CPU voltage margins. We introduce a run-time governor that dynamically reduces the supply voltage of modern multicore Intel x86-64 CPUs. Our governor employs a model that takes as input a set of performance metrics which are directly measurable via performance monitoring counters and have high predictive value for the minimum tolerable supply voltage, to dynamically predict and apply the appropriate reduction for the workload at hand. Compared to the conventional DVFS governor of Intel x86-64 CPUs, our approach achieves significant energy savings, up to 42% and 34%, respectively, for two off-the-shelf processor families. The quantification and exploitation of CPU voltage margins requires long experimental campaigns, involving multiple systems, multiple configurations and numerous different workloads. Moreover, these experimental campaigns have different targets: characterization of CPU voltage margins, data collection, validation and experimental evaluation of the proposed techniques. Motivated by the complexity of these experimental campaigns, we developed a framework that significantly simplifies the configuration and automates the execution of such experimental campaigns. It supports multiple execution modes, namely OS-controlled and bare-metal execution, it can recover from system crashes due to aggressively reduced voltage margins, and can detect symptoms of erratic workload behavior in the form of Machine Check Exceptions (MCEs), system log entries, or even Silent Data Corruptions (SDCs) if the correct output is supplied as a reference. Finally, we investigate whether the combination of heterogeneous and approximate computing can yield favorable solutions in the energy efficiency vs. quality of results tradeoff. Approximate computing minimizes the energy footprint of applications at the expense of output quality. More specifically, under the premise that not all parts or execution phases of a program affect the quality of its output equally, developers can introduce approximations in less significant parts of their code in an educated manner, in order to improve the energy efficiency of code execution. We experiment with a set of applications from different domains, with diverse characteristics. We have modified the applications to exploit both heterogeneity and approximations. We evaluate them on heterogeneous platforms (comprising of CPUs and GPUs) and quantify the isolated and combined effect of heterogeneous and approximate computing. Our results show that heterogeneous and approximate computing, independently, result in significant energy gains. Moreover, they can be combined to drastically minimize the energy footprint of executions by up to 94%.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2020

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Koutsovasilis, Panagiotis
  • Κουτσοβασίλης, Παναγιώτης

Subjects

dc:subject × 14

Rights

dc:rights
Statement dc:rights
  • BY
Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/47334

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Koutsovasilis, Panagiotis; Κουτσοβασίλης, Παναγιώτης. Exploiting intrinsic hardware guardbands and software heterogeneity to improve the energy efficiency of computing systems. University of Thessaly (UTH), 2020. http://hdl.handle.net/10442/hedi/47334