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%.

Emerging applications in healthcare, autonomous vehicles, and wearable assistance require interactive and low-latency data analysis services. Unfortunately, cloud-centric architectures cannot fulfill the low-latency demands of these applications, as user devices are often distant from cloud data centers. Edge computing aims to reduce the latency by enabling processing tasks to be offloaded to resources located at the network’s edge. However, determining which tasks must be offloaded to edge servers to reduce the latency of application requests is not trivial, especially if the tasks present dependencies. This paper proposes a Deep Reinforcement Learning (DRL) approach called TPTO, which leverages Transformer Networks and Proximal Policy Optimization (PPO) to offload dependent tasks of IoT applications in edge computing. We consider users with various preferences, where devices can offload computation to an edge server via wireless channels. Performance evaluation results demonstrate that under fat application graphs, TPTO is more effective than state-of-the-art methods, such as Greedy, HEFT, and MRLCO, by reducing latency by 30.24%, 29.61%, and 12.41%, respectively. In addition, TPTO presents a training time approximately 2.5 times faster than an existing DRL approach.

Internet of Things (IoT) applications often require the processing of data streams generated by devices dispersed over a large geographical area. Traditionally, these data streams are forwarded to a distant cloud for processing, thus resulting in high application end-to-end latency. Recent work explores the combination of resources located in clouds and at the edges of the Internet, called cloud-edge infrastructure, for deploying Data Stream Processing (DSP) applications. Most previous work, however, fails to scale to very large IoT settings. This paper introduces deployment strategies for the placement of Data Stream Processing (DSP) applications onto cloud-edge infrastructure. The strategies split an application graph into regions and consider regions with stringent time requirements for edge placement. The proposed Aggregate End-to-End Latency Strategy with Region Patterns and Latency Awareness (AELS+RP+LA) decreases the number of evaluated resources when computing an operator's placement by considering the communication overhead across computing resources. Simulation results show that, unlike the state-of-the-art, Aggregate End-to-End Latency Strategy with Region Patterns and Latency Awareness (AELS+RP+LA) scales to environments with more than 100k resources with negligible impact on the application end-to-end latency.

The Internet of Things has enabled many application scenarios where a large number of connected devices generate unbounded streams of data, often processed by data stream processing frameworks deployed in the cloud. Edge computing enables offloading processing from the cloud and placing it close to where the data is generated, thereby reducing the time to process data events and deployment costs. However, edge resources are more computationally constrained than their cloud counterparts, raising two interrelated issues, namely deciding on the parallelism of processing tasks (a.k.a. operators) and their mapping onto available resources. In this work, we formulate the scenario of operator placement and parallelism as an optimal mixed-integer linear programming problem. The proposed model is termed as Cloud-Edge data Stream Placement (CESP). Experimental results using discrete-event simulation demonstrate that CESP can achieve an end-to-end latency at least ≃ 80% and monetary costs at least ≃ 30% better than traditional cloud deployment.

The number of Internet of Things applications is forecast to grow exponentially within the coming decade. Owners of such applications strive to make predictions from large streams of complex input in near real time. Cloud-based architectures often centralize storage and processing, generating high data movement overheads that penalize real-time applications. Edge and Cloud architecture pushes computation closer to where the data is generated, reducing the cost of data movements and improving the application response time. The heterogeneity among the edge devices and cloud servers introduces an important challenge for deciding how to split and orchestrate the IoT applications across the edge and the cloud. In this paper, we extend our IoT Edge Framework, called R-Pulsar, to propose a solution on how to split IoT applications dynamically across the edge and the cloud, allowing us to improve performance metrics such as end-to-end latency (response time), bandwidth consumption, and edge-to-cloud and cloud-to-edge messaging cost. Our approach consists of a programming model and real-world implementation of an IoT application. The results show that our approach can minimize the end-to-end latency by at least 38% by pushing part of the IoT application to the edge. Meanwhile, the edge-to-cloud data transfers are reduced by at least 38% and the messaging costs are reduced by at least 50% when using the existing commercial edge cloud cost models.

There is increasing demand for handling massive amounts of data in a timely manner via Distributed Stream Processing (DSP). A DSP application is often structured as a directed graph whose vertices are operators that perform transformations over the incoming data and edges representing the data streams between operators. DSP applications are traditionally deployed on the Cloud in order to explore the virtually unlimited number of resources. Edge computing has emerged as a suitable paradigm for executing parts of DSP applications by offloading certain operators from the Cloud and placing them close to where the data is generated, hence minimising the overall time required to process data events (i.e., the end-to-end latency). The operator reconfiguration consists of changing the initial placement by reassigning operators to different devices given target performance metrics. In this work, we model the operator reconfiguration as a Reinforcement Learning (RL) problem and define a multi-objective reward considering metrics regarding operator reconfiguration, and infrastructure and application improvement. Experimental results show that reconfiguration algorithms that minimise only end-to-end processing latency can have a substantial impact on WAN traffic and communication cost. The results also demonstrate that when reconfiguring operators, RL algorithms improve by over 50% the performance of the initial placement provided by state-of-the-art approaches.

Distributed Stream Processing (DSP) applications are increasingly used in new pervasive services that process enormous amounts of data in a seamless and near real-time fashion. Edge computing has emerged as a means to minimise the time to handle events by enabling processing (i.e., operators) to be offloaded from the Cloud to the edges of the Internet, where the data is often generated. Deciding where to execute such operations (i.e., edge or cloud) during application deployment or at runtime is not a trivial problem. In this work, we employ Reinforcement Learning (RL) and Monte-Carlo Tree Search (MCTS) to reassign operators during application runtime. Experimental results show that RL and MCTS algorithms perform better than traditional placement techniques. We also introduce an optimisation to a MCTS algorithm, called MCTS-Best-UCT, that achieves similar latency with fewer operator migrations and faster execution time. In certain scenarios, the time needed by MCTS-Best-UCT to find the best end-to-end latency is at least 33% smaller than the time required by the other algorithms.

Allocating IT resources to Virtual Infrastructures (VIs) (i.e.groups of VMs, virtual switches, and their network interconnections) is an NP-hard problem. Most allocation algorithms designed to run on CPUs face scalability issues when considering current cloud data centers comprising thousands of servers. This work offers and evaluates a set of allocation algorithms refactored for Graphic Processing Units (GPUs). Experimental results demonstrate their ability to handle three large-scale data center topologies.

Data Stream Processing (DSP) is a widely used programming paradigm to process an unbounded event stream. Often, DSP frameworks are deployed on the cloud with a scalable resource model. One of the key requirements of DSP is to produce results with low latency. With the emergence of IoT, many event sources have been located outside the cloud which can result in higher end-to-end latency due to communication overhead. However, due to the abundance of resources at the IoT layer, Edge computing has emerged as a viable computational paradigm. In this paper, we devise an optimisation framework, consisting of a constraint satisfaction formulation and a system model, that aims to minimise end-to-end latency through appropriate placement of DSP operators either on cloud nodes or edge devices, i.e. deployed in an edge-cloud integrated environment. We test our optimisation framework using OMNeT++, with realistic topologies and power consumption data, and show that it is capable of achieving approx 1.65 times reduction of latency compared to edge-only and cloud-only placements, which in turn also reduces the energy consumption per event by up to approx 4% at the edge layer. To the best of our knowledge our optimisation framework is the first of its kind to integrate power, bandwidth and CPU constraints with latency minimisation.

Traffic engineering technologies such as MPLS have been proposed to adjust the paths of data flows according to network availability. Although the time interval between traffic optimisations is often on the scale of hours or minutes, modern SDN techniques enable reconfiguring the network more frequently. It is argued, however, that changing the paths of TCP flows too often could severely impact their performance by incurring packet loss and reordering. This work analyses and evaluates the impact of frequent route changes on the performance of TCP flows. Experiments carried out on a network testbed show that rerouting a flow can affect its throughput when reassigning it a path either longer or shorter than the original path. Packet reordering has a negligible impact when compared to the increase of RTT. Moreover, constant rerouting influences the performance of the congestion control algorithm. Designed to assess the limits on SDN-induced reconfiguration, a scenario where the traffic is rerouted every 0.1s demonstrates that the throughput can be as low as 35% of that achieved without rerouting.

This work investigates factors that can impact the elasticity of bare-metal resources. We analyse data from a real bare-metal deployment system to build a deployment time model, which is used to evaluate how provisioning time impacts the reservation of bare-metal resources. Climate/Blazar, a reservation framework designed for OpenStack, is discussed. Simulation results show that reservations can help reduce the time to deliver a provisioned cluster to its customer while achieving energy savings similar to those of strategies that switch-off idle resources.

Reducing the energy consumed by wired computer networks is a challenge that has been actively investigated over the past few years. A popular mechanism proposed to reduce the consumption aims to put links and line cards to sleep mode during off-peak hours. Such a mechanism, however, decreases the available network capacity and increases the risk of congestion if traffic rises unexpectedly. This paper proposes a solution to rapidly react to network bursts and turn-on sleeping links, which we term as SegmenT Routing based Energy Efficient Traffic Engineering for switching ON (STREETE-ON). The proposed algorithm was implemented in the OMNeT++ network simulator using state-of-art dynamic graph algorithms. In such a way, we achieved execution times of tens of milliseconds for a 50-node network. Experimental results show that STREETE-ON can effectively prevent network congestion, avoid turning-on unneeded links, and preserve good energy-efficiency of the network.

An important feature of cloud computing is its elasticity, that is, the ability to have resource capacity dynamically modified according to the current system load. Auto-scaling is challenging because it must account for two conflicting objectives: minimising system capacity available to users and maximising QoS, which typically translates to short response times. Current auto-scaling techniques are based solely on load forecasts and ignore the perception that users have from cloud services. As a consequence, providers tend to provision a volume of resources that is significantly larger than necessary to keep users satisfied. In this article, we propose a scheduling algorithm and an auto-scaling triggering technique that explore user patience in order to identify critical times when auto-scaling is needed and the appropriate volume of capacity by which the cloud platform should either extend or shrink. The proposed technique assists service providers in reducing costs related to resource allocation while keeping the same QoS to users. Our experiments show that it is possible to reduce resource-hour by up to approximately 8% compared to auto-scaling based on system utilisation.

Auto-scaling is a key feature in clouds responsible for adjusting the number of available resources to meet service demand. Resource pool modifications are necessary to keep performance indicators, such as utilisation level, between user-defined lower and upper bounds. Auto-scaling strategies that are not properly configured according to user workload characteristics may lead to unacceptable QoS and large resource waste. As a consequence, there is a need for a deeper understanding of auto-scaling strategies and how they should be configured to minimise these problems. In this work, we evaluate various auto-scaling strategies using log traces from a production Google data centre cluster comprising millions of jobs. Using utilisation level as performance indicator, our results show that proper management of auto-scaling parameters reduces the difference between the target utilisation interval and the actual values-we define such difference as Auto-scaling Demand Index. We also present a set of lessons from this study to help cloud providers build recommender systems for auto-scaling operations.

Applications for mobile devices are increasingly relying on Cloud services to provide content and offload data processing tasks. Traditionally, web-based systems have been optimised to improve response time. Touch sensitive screens and the various sensors of mobile devices allow for better instrumentation, enabling providers to obtain more honest signals on how users utilise a service and learn about their behaviours. This work introduces an architecture that explores honest signals to determine a few user behaviours, e.g. the tendency to perform multiple tasks at a time, change focus, and expect fast response from a service. The architecture relies on a novel resource management strategy that considers such behaviours to prioritise service requests from users who demand fast response from a service. By comparing the proposed strategy with one that does not consider the signals, we show that the experience of users who demand faster response time can be improved without degrading the quality of service of those who often perform multiple activities. The proposed strategy also brings benefits to service providers as no additional resources are necessary to enhance overall user experience, which we modelled using Prospect Theory.

Increasingly organizations are considering moving their workloads to clouds to take advantage of the anticipated benefits of a more cost effective and agile IT infrastructure. A key component of a cloud service, as it is exposed to the consumer, is the published selection of instance resource configurations (CPU, memory, and disk). The number of instance configurations, as well as the specific values that characterize them, form important decisions for the cloud service provider. This paper explores these resource configurations; examines how well a traditional data center fits into the cloud model from a resource allocation perspective; and proposes a framework, named CloudAffinity, aimed at selecting an optimal number of configurations based on customer requirements.

In this paper, we investigate the benefits that organisations can reap by using "Cloud Computing" providers to augment the computing capacity of their local infrastructure. We evaluate the cost of six scheduling strategies used by an organisation that operates a cluster managed by virtual machine technology and seeks to utilise resources from a remote Infrastructure as a Service (IaaS) provider to reduce the response time of its user requests. Requests for virtual machines are submitted to the organisation's cluster, but additional virtual machines are instantiated in the remote provider and added to the local cluster when there are insufficient resources to serve the users' requests. Naïve scheduling strategies can have a great impact on the amount paid by the organisation for using the remote resources, potentially increasing the overall cost with the use of IaaS. Therefore, in this work we investigate six scheduling strategies that consider the use of resources from the "Cloud", to understand how these strategies achieve a balance between performance and usage cost, and how much they improve the requests' response times.

Numerous Grids have been created during the last years. Most of these Grids work in isolation and with different utilisation levels. As the resource utilisation within a Grid has fixed and operational costs, there can be benefits for a Grid to offload requests to another Grid or provide spare resources, thus reducing the cost of over-provisioning. In this work, we enable load management across Grids through resource exchange between them considering the cost for one Grid to acquire resources from another. However, enabling resource exchange amongst Grids is a challenging task: a Grid should not compromise the performance of its local user communities' applications, yet benefit from providing spare resources to other Grids. The load management mechanism and related policies take into consideration the economic compensation of providers for the resources allocated. Experimental results show that the mechanism achieves its goal in redirecting requests, increasing the number of user requests accepted and balancing the load amongst Grids.

Abstract Over the last few years, several nations around the world have set up Grids to share resources such as computers, data, and instruments to enable collaborative science, engineering, and business applications. These Grids follow a restricted organizational model wherein a Virtual Organization (VO) is created for a specific collaboration and all interactions such as resource sharing are limited to within the VO. Therefore, dispersed Grid initiatives have led to the creation of disparate Grids with little or no interaction between them. In this paper, we propose a model that: (a) promotes interlinking of islands of Grids through peering arrangements to enable InterGrid resource sharing; (b) provides a scalable structure for Grids that allow them to interconnect with one another and grow in a sustainable way; (c) creates a global Cyberinfrastructure to support e‐Science and e‐Business applications. This work identifies and proposes architecture, mechanisms, and policies that allow the internetworking of Grids and allows Grids to grow in a similar manner as the Internet. We term the structure resulting from such internetworking between Grids as theInterGrid. The proposed InterGrid architecture is composed of InterGrid Gateways responsible for managing peering arrangements between Grids. We discuss the main components of the architecture and present a research agenda to enable the InterGrid vision. Copyright © 2007 John Wiley & Sons, Ltd.

Intrusion detection systems (IDSs) have been substantially improved in recent past. However, network attacks have become more sophisticated and increasingly complex: many of current attacks are coordinated and originated in multiple networks. To detect these attacks, IDSs need to obtain information on network events from multiple networks or administrative domains. This work demonstrates that a Distributed IDS (DIDS) can be composed of existing IDSs, improving the detection of misuses in a multiple network environment. We use a grid middleware for creating a service-based intrusion detection grid. We demonstrate through experimental results that the proposed DIDS allows the integration of heterogeneous existing IDSs and improves the detection of attacks by exploring the synergy between existing IDSs.

In this article, we describe a grid of sensors to collect patients' vital data and to allow real time monitoring of patients in heath-care centres. We analyse the problem scenario and identify the components involved towards the construction of an integrated and homogeneous management system. Finally, we present a case study to demonstrate the applicability of our approach

In this paper we present a proposal of a middleware to integrate mobile computing with grid computing. We describe the support to be provided by the combination of the two technologies, propose a middleware to support the development and management of mobile grids, and discuss how this scenario contributes to the development of a new age of mobile service applications

Abstract The centralized system approach for computer and telecommunication network management has been presenting scalability problems along with the growth in the amount and diversity of managed equipment. Moreover, the increase in complexity of the services being offered through the networks also contributes to adding extra workload to the management station. The amount of data that must be handled and processed by only one administration point could lead to a situation where there is not enough processing and storage power to carry out an efficient job. In this work we present an alternative approach by creating a highly distributed computing environment through the use of Grids of autonomous agents to analyze large amounts of data, which reduce the processing costs by optimizing the load distribution and resource utilization. Copyright © 2004 John Wiley & Sons, Ltd.