Search

One of the largest single uses of energy, and in most countries therefore oil and gas, is the heating and cooling of buildings. Much of the built environment we need to decarbonise by 2050 is already built. Adapting this infrastructure is going to be disruptive and expensive, and take time we arguably do not have. The main current approach to limiting this consumption is to try to tightly control the indoor conditions to a given setpoint temperature. But, is there another approach? We have conducted a year-long study of the actual thermal performance of a large non-domestic building. We find that there are three significant socio-technical gaps between the buildings systems’ perception of how the building performs and the temperatures experienced by building users which affect peoples’ comfort and waste energy. In contrast with traditional approaches, we argue for more flexibility and adaptivity in both the policy and and how the building is controlled to address this. We believe that mitigating these gaps and avoiding the wasted energy and associated indirect emissions is a significant opportunity for future ICT for sustainability systems.

Deep Learning (DL) models are deployed as jobs within machines containing GPUs. These DL systems - ranging from a singular GPU device to machine clusters - require state-of-the-art resource management to increase resource utilization and job throughput. While it has been identified that co-location - multiple jobs co-located within the same GPU - is an effective means to achieve this, such co-location incurs performance interference that directly debilitates DL training and inference performance. Existing approaches to mitigate interference require resource intensive and time consuming kernel profiling ill-suited for runtime scheduling decisions. Current DL system resource management are not designed to deal with these problems. This paper proposes Horus, an interference-aware resource manager for DL systems. Instead of leveraging expensive kernel-profiling, our approach estimates job resource utilization and co-location patterns to determine effective DL job placement to minimize likelihood of interference, as well as improve system resource utilization and makespan. Our analysis shows that interference cause up to 3.2x DL job slowdown. We integrated our approach within the Kubernetes resource manager, and conduct experiments in a DL cluster by training 2,500 DL jobs using 13 different models types. Results demonstrate that Horus is able to outperform other DL resource managers by up to 61.5% for resource utilization and 33.6% for makespan.

Deep Learning (DL) models are deployed as jobs within machines containing GPUs. These DL systems - ranging from a singular GPU device to machine clusters - require state-of-the-art resource management to increase resource utilization and job throughput. While it has been identified that co-location - multiple jobs co-located within the same GPU - is an effective means to achieve this, such co-location incurs performance interference that directly debilitates DL training and inference performance. Existing approaches to mitigate interference require resource intensive and time consuming kernel profiling ill-suited for runtime scheduling decisions. Current DL system resource management are not designed to deal with these problems. This paper proposes Horus, an interference-aware resource manager for DL systems. Instead of leveraging expensive kernel-profiling, our approach estimates job resource utilization and co-location patterns to determine effective DL job placement to minimize likelihood of interference, as well as improve system resource utilization and makespan. Our analysis shows that interference cause up to 3.2x DL job slowdown. We integrated our approach within the Kubernetes resource manager, and conduct experiments in a DL cluster by training 2,500 DL jobs using 13 different models types. Results demonstrate that Horus is able to outperform other DL resource managers by up to 61.5% for resource utilization and 33.6% for makespan.

Designing with the community brings about a number of benefits, including tacit and contextual knowledge about the problem domain; this is especially apparent in rural settings. However, designing for problems that have yet to embed themselves in the fabric of society (i.e. future problems) poses a number of challenges, as they typically present intangible scenarios and concepts that have yet to be experienced by the wider-community. Using the OnSupply project as a case study, we share our experience in working with the Tiree community to address a future problem through a technology-mediated enquiry. Furthermore, we present a novel process that uses creative workshops augmented with physical artefacts to inform and learn from the community about a problem space, and to seed the design of a system that addresses it.

One of the key challenges faced when developing context-aware pervasive systems is to capture the set of inputs that we want a system to adapt to. Arbitrarily specifying ranges of sensor values to respond to will lead to incompleteness of the specification, and may also result in conflicts, when multiple incompatible adaptations may be triggered by a single user action. We posit that the ideal approach combines the use of past traces of real, annotated context data with the ability for a system designer or user to go in and interactively modify the specification of the set of inputs a particular adaptation should be responsive to. We introduce Situvis, an interactive visualisation tool we have developed which assists users and developers of context-aware pervasive systems by visually representing the conditions that need to be present for a situation to be triggered in terms of the real-world context that is being recorded, and allows the user to visually inspect these properties, evaluate their correctness, and change them as required. This tool provides the means to understand the scope of any adaptation defined in the system, and intuitively resolve conflicts inherent in the specification.

This paper presents an analysis, implementation and evaluation of the physical mobile interaction techniques touching, pointing and scanning. Based on this we have formulated guidelines that show in which context which interaction technique is preferred by the user. Our main goal was to identify typical situations and scenarios in which the different techniques might be useful or not. In support of these aims we have developed and evaluated, within a user study, a low-fidelity and a high-fidelity prototype to assess scanning, pointing and touching interaction techniques within different contexts. Other work has shown that mobile devices can act as universal remote controls for interaction with smart objects but, to date, there has been no research which has analyzed when a given mobile interaction technique should be used. In this research we analyze the appropriateness of three interaction techniques as selection techniques in smart environments.

This paper is about our experiences of space in the Equator project (www.equator.ac.uk), in particular, the way in which multiple spaces, both virtual and physical, can co-exist. By this we mean that people and objects may have locations in and relationships to both physical space and one or more virtual spaces, and that these different spaces together interact to give an overall system behaviour and user experience. The concepts we develop in this chapter are driven partly by practical experience, and partly by previous theoretical work such as the models and taxonomies of spatial context in (Dix et al., 2000), the models for mixed reality boundaries (Koleva et al., 1999) and capturing human spatial understanding exposed in sources such maps, myths and mathematics (Dix, 2000). We are also building on established work on informal reasoning about space from the AI and GIS communities (Grigni et al, 1995; Papadias et al., 1996) similar to Allen’s well known temporal relations (Allen, 1991).

Mobile applications are required to operate in environments characterised by change. More specifically, the availability of resources and services may change significantly during a typical period of system operation. As a consequence, adaptive mobile applications need to be capable of adapting to these changes to ensure they offer the best possible level of service to the user. Our experiences of developing and evaluating adaptive context-aware applications in mobile environments has led us to believe that existing architectures fail to provide the necessary support for such applications. In this paper, we discuss the shortcomings of existing approaches and present work on our own architecture that has been designed to meet the key requirements of context-aware adaptive applications.

This paper introduces a novel access control architecture for publicly accessible, wireless networks. The architecture was designed to address the requirements obtained from a case study of ubiquitous Internet service provisioning within the city of Lancaster. The proposed access control mechanism is. based on the concepts of secure user authentication, packet marking, and packet filtering at the access routers. The paper demonstrates to what extent this token-based, soft-state access control mechanism improves security and robustness, and offers improved performance over that provided by existing approaches within roaming networks. Early indications show the access control mechanism can better be implemented through the use of active routers, in order to facilitate dynamic rollout and configuration of the system. In addition, extensions to Mobile IPv6 are proposed, which provide support for roaming users at a fundamental level.