Back to search

University of Thessaly (UTH)

Software policies and techniques for improving the energy and cost efficiency of cloud and high-performance computing infrastructures

Abstract

dc:description

Energy consumption is a critical concern in the areas of Cloud- and High-Performance Computing (HPC) due to economic and environmental reasons. To fulfill the requirements of the compute jobs they serve, modern datacentres are equipped with multi-core processors and high amounts of Dynamic Random Access Memory (DRAM), which significantly increase the energy footprint of the nodes. Therefore, the minimization of energy consumption is a challenging undertaking, as infrastructure providers need, at the same time, to meet the quality of service (QoS) requirements of the running services and applications. Cloud infrastructure providers are contractually bound with Service Level Agreements (SLAs) with their users, and the violation of SLAs translates to penalties and thus cost. There are several mechanisms/techniques that can reduce the energy consumption of the datacentres, while also considering the QoS. VM consolidation targets to pack as many VMs as possible on the nodes, ideally without performance loss, in order to reduce the number of active nodes. Heterogeneity can also improve the energy efficiency of the nodes; accelerators such as GPUs are superior in terms of performance per Watt compared with the conventional CPUs. If heterogeneity is combined with approximate computing the energy gains can be even higher, as the relaxation of some calculations may introduce more energy gains and favor certain architectures. On the hardware side, manufacturers have introduced Voltage and Frequency Scaling (VFS), enabling the processor to operate at more power-efficient operating points, whenever there is an opportunity to do so without penalizing the performance. However, at the same time, CPU and DRAM manufacturers introduce guardbands to the operating parameters of the processors (supply voltage for each frequency step) and the DRAMs (supply voltage and refresh rate) to guarantee correct operation under adverse combinations of operating conditions. Those guardbands are unnecessarily wide for real-world execution scenarios and translate to increased power consumption. In this Thesis, we focus on improving the energy efficiency of cloud infrastructures by reducing hardware guardbands. We formally model the problem of dealing with the trade-off between energy cost and potential SLA violations by considering the VM requirements and the fact that they change over time. Moreover, we introduce scheduling policies that use these models and exploit reduced margins for the CPU and the DRAMs to lower the energy cost of cloud datacentres. Finally, we combine heterogeneity with approximate computing in HPC applications, and we evaluate the trade-off between the quality of output and the energy footprint of these applications. More specifically: We model the cost for cloud/edge infrastructure providers studying the effects of the conventional technique of VFS and the more aggressive approaches of voltage (undervolting) and frequency (overclocking) overscaling of the processor. We introduce models for estimating the failure rate for each of the aforementioned techniques (VFS, undervolting, overclocking). Our models capture the expected energy consumption of the tasks as well as potential penalties due to SLA violations. Our study shows up to 13.31% cost reduction in the case of undervolting compared with the operation of the processor at nominal settings. As a next step, we use a simulated environment to investigate at-scale the exploitation of reduced voltage guardbands on the processor by cloud infrastructure providers. We introduce a VM scheduling and node configuration policy that exploits operation at reduced voltage margins for the processor, on top of the combination of the conventional techniques of VFS and VM consolidation. The policy is based on an analytic and architecture-agnostic model that captures the trade-off between energy cost and potential SLA violation penalties. We evaluate our approach and compare it against a state-of-the-art scheduling and configuration policy. We use node profiles derived from the characterization of commercially available server-class systems. Moreover, for our experiments we use realistic values for the energy price and the price of VMs. We provide results with both a constant and a dynamically varying energy price. Our results show gains up to 17.24% and 17.72% for the cost and the energy, respectively, compared with a state-of-art policy that combines VFS with VM consolidation. We also explore the trade-off between energy efficiency and node overloads, to manage VM consolidation in a cost-aware manner. More specifically, we introduce a model that uses runtime information to control the load of the nodes, identifying and considering correlated behavior among collocated workloads. As the requirements of VMs are not known in advance, it is critical to consider recent VM behavior to adapt the target load for each node. We combine this approach with reduced voltage margins for the processor and VFS. We show that our policy closely matches, or even outperforms two state-of-the-art policies that combine VM consolidation with VFS – the second one, also operating the CPU at reduced voltage margins. A key differentiation with prior policies is that our method does not require an unrealistic, in-advance analysis of the characteristics of the workload. We also study the exploitation of reduced margins for the operating parameters of DRAMs, namely the supply voltage and refresh rate. In contrast with CPUs, where an error mainly translates to a system crash, for DRAMs we study the effect of potential penalties due to correctable single- and detectable double-bit errors. In analogy with the work on CPUs, we introduce an architecture-agnostic model and a novel workload management and system configuration policy. Our model captures all the aspects of VM management and DRAM configuration. The policy we introduce uses this model to exploit reduced margins of the operating parameters of DRAMs, combined with VM consolidation. As future systems will be equipped with high DRAM capacities, we study this problem for different memory sizes. We, again, evaluate our approach at-scale on a simulated environment, using realistic node parameters and realistic values for the energy price and VM pricing. Our results show that 29.53% and 34.84% cost and energy gains can be achieved, compared with a state-of-art policy that exploits VM consolidation. The gains are amplified as we move to nodes with a higher memory capacity. Finally, we examine the effect of heterogeneous and approximate computing on the energy footprint of HPC applications. Result quality is not equally affected by all computations of a program; therefore we can perform selected computations approximately to improve the energy efficiency without heavily penalizing the output quality. In our work, we use 4 applications from diverse domains, ranging from real-world applications to kernels. We modify these applications to exploit both heterogeneity and approximate computing. We evaluate our research on heterogeneous systems (equipped with CPUs and GPUs), and we quantify the benefits of heterogeneity and approximate computing, considering the trade-off between energy consumption and the quality of the output. Our results show that heterogeneity and approximate computing can offer significant energy gains.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2024

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kalogirou, Christos
  • Καλογήρου, Χρήστος

Subjects

dc:subject × 30

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/56569

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

Kalogirou, Christos; Καλογήρου, Χρήστος. Software policies and techniques for improving the energy and cost efficiency of cloud and high-performance computing infrastructures. University of Thessaly (UTH), 2024. http://hdl.handle.net/10442/hedi/56569