As the demand for AI-driven workloads increases, the energy consumption of Graphics Processing Units (GPUs) devices has come under intense scrutiny, particularly in hyperscale data centers where large numbers of accelerators are centralized and leased to diverse clients.

International audience

Energy consumption has become a critical concern in Information and Communication Technologies (ICT), pressing for more accurate measurements. While the power consumption of physical servers can be physically monitored, organizations are increasingly adopting virtual environments, such as cloud computing, rendering physical measurements impractical in operational contexts. The state-of-the-art approaches to estimating this ”virtual” consumption mostly consist of assigning server power consumption shares among hosted processes, guided by various system metrics. Unfortunately, such a bottom-up approach is highly sensitive in a multi-tenant environment, thus failing to report stable measurements to stakeholders. For example, the same activity performed by one Virtual Machine (VM) may lead to different power consumption traces, depending on the activity of the co-hosted VMs. As cloud customers have only control over their provisioned virtual resources, we propose a new method to model the power consumption of their virtual appliances, enabling contextagnostic tracking of their environmental impact. This framework, called CINERGY, is designed to be more predictable than the state-of-the-art power models, while still exposing the gains from consolidation. We evaluate its accuracy against ground-truth measurements, often lacking in the literature. We show that CINERGY is deterministic and accurate, with an average error of 6.6%.

The adoption of computing resources oversubscription in cloud environments is conventionally limited to a restricted subset of Virtual Machines (VMs) within the providers’ offerings, primarily driven by performance considerations. So far, VMs schedulers mostly implement all-or-nothing oversubscription strategies, wherein all VM resources are either oversubscribed or remain unaltered. While the former strategy offers higher consolidation rates, the latter delivers better performance guarantees.In this paper, we conducted an empirical study of the individual usage of virtual CPUs (vCPUs) in the OVHCloud production environment and we demonstrate that, as they are not uniformly utilized, the current holistic approach may not be appropriate. Based on these observations, we introduce a novel approach, named SweetspotVM, where oversubscription ratios are applied at the granularity of individual vCPU, instead of the whole VMs. This novel paradigm unlocks a more flexible oversubscription management strategy, pinning oversubscription ratios per vCPU within VMs. We present a prototype of SweetspotVM to illustrate the feasibility of accommodating multiple oversubscription levels within a single host and assigning them to individual vCPU.We assess the viability of our approach on a physical platform, demonstrating the possibility of dividing the cost of hosting VMs by 3, while maintaining the VMs performance at the level of non-oversubscribed platforms. We, therefore, believe that SweetspotVM opens new avenues to boost the consolidation of VMs on a reduced number of servers, with positive impacts on the environmental footprint of cloud computing.

Cloud providers generally expose a large catalog of Virtual Machine (VM) offers, some being categorized as premium-guaranteeing dedicated resources-and others being hosted in oversubscribed environments, where virtual resources can exceed the physical capabilities of Physical Machines (PMs). The latter strategy is often employed to increase platform utilization, as hosted VMs are unlikely to fully utilize all their allocated resources simultaneously [1]. However, managing multiple oversubscribed VM levels introduces an additional layer of complexity for Cloud providers, often leading them to provision isolated clusters of PMs for each category of offers. In this paper, we introduce SLACKVM, a novel Cloud-shared architecture wherein VMs from various oversubscription levels coexist on the same cluster of PMs. In particular, we demonstrate that oversubscription levels can be complementary, meaning they do not saturate the same resource components. By leveraging this complementarity, Cloud providers can couple multiple levels to better consolidate VM offers onto PMs, and reduce the size of their clusters by up to 9.6%. These resource savings result in both an operational cost reduction and a reduced ecological footprint for Cloud infrastructures, with a limited impact on the Quality of Service (QoS).

Despite continuous improvements, cloud physical resources remain underused, hence severely impacting the efficiency of these infrastructures at large. To overcome this inefficiency, Infrastructure-as-a-Service (IaaS) providers usually compensate for oversized Virtual Machines (VMs) by offering more virtual resources than are physically available on a host. However, this technique—known asoversubscription—may hinder performances when a statically-defined oversubscription ratio results in resource contention of hosted VMs. Therefore, instead of setting a static and cluster-wide ratio, this article studies how a greedy increase of the oversubscription ratio per Physical Machine (PM) and resources type can preserve performance goals. Keeping performance unchanged allows our contribution to be more realistically adopted by production-scale IaaS infrastructures. This contribution, namedScroogeVM, leverages the detection of PM stability to carefully increase the associated oversubscription ratios. Based on metrics shared by public cloud providers, we investigate the impact of resource oversubscription on performance degradation. Subsequently, we conduct a comparative analysis ofScroogeVMwith state-of-the-art oversubscription computations. The results demonstrate that our approach outperforms existing methods by leveraging the presence of long-lasting VMs, while avoiding live migration penalties and performance impacts for stakeholders.

Cloud infrastructures are large-scale and complex platforms designed to host a wide diversity of applications and workloads. Given these complexity and scale factors, simulators and benchmarks are broadly adopted in vitro to study their behaviors, prototype new software components and heuristics, and evaluate their effective performances.However, both state-of-the-art simulations and benchmarks may suffer from a representativeness problem, as the reported results can vary depending on their input workloads. For example, a Infrastructure-as-a-Service (IaaS) platform aims to host Virtual Machines (VMs), whose characteristics (resource configurations, workload intensity, arrival/departure rate, etc.) can greatly differ depending on Cloud providers and public/private deployments. Addressing this IaaS representativeness thus requires Cloud providers to share production-scale datasets, which might be considered sensitive. Moreover, Simulations and benchmarks require a specific experiment scenario that cannot be easily generated from Cloud providers characteristics.To address these issues, this paper introduces CloudFactory, a IaaS workload generator. Our contribution is first composed of a library that can be used by Cloud providers to share IaaS statistics, instead of raw datasets. Then, we introduce a generator designed to produce realistic VM workloads that match these statistics. CloudFactory is made available as open-source software that can be adopted by Cloud providers and researchers to foster the evaluation of new contributions.As an example, we perform an analysis on scheduling evolution for different IaaS workload intensity of two different Cloud providers: Microsoft Azure and Chameleon. We also report on OVHcloud statistics computed from CloudFactory and compare them to other Cloud providers.