The task of knowledge tracing involves establishing a model that characterizes the mastery level of individual students in relation to knowledge concepts (KCs), as they engage with a sequence of learning activities. Each student’s knowledge is modelled by estimating the student’s performance on the learning activities. This domain of research holds significant importance in the development of personalized learning platforms for students. However, within the current landscape, the majority of existing Knowledge Tracing (KT) methods exhibit a gap in their approach, as they often fail to incorporate two pivotal factors simultaneously: the in-dividualization of student characteristics (i.e., the prior knowledge and learning rates) and the interconnectedness between KCs. In essence, these existing approaches tend to disregard the intrinsic variations in students’ prior knowledge and learning paces, which invariably differ from one student to another. In this study, our focus lies in predicting student performance through modelling his/her knowledge state, while also extracting dependencies between KCs and his/her prior learning interactions. To this end, we propose a novel self Attentive Convolutional Knowledge Tracing (ACKT) method for analyzing continuous learning interactions of students. Extensive experiments on several real-world benchmark datasets show that ACKT could obtain better knowledge tracing results and it outperforms existing KT methods for predicting future learner responses.

Window aggregation queries are a core part of streaming applications. To support window aggregation efficiently, stream processing engines face a trade-off between exploiting parallelism (at the instruction/multi-core levels) and incremental computation (across overlapping windows and queries). Existing engines implement ad-hoc aggregation and parallelization strategies. As a result, they only achieve high performance for specific queries depending on the window definition and the type of aggregation function. We describe a general model for the design space of window aggregation strategies. Based on this, we introduce LightSaber, a new stream processing engine that balances parallelism and incremental processing when executing window aggregation queries on multi-core CPUs. Its design generalizes existing approaches: (i) for parallel processing, LightSaber constructs a parallel aggregation tree (PAT) that exploits the parallelism of modern processors. The PAT divides window aggregation into intermediate steps that enable the efficient use of both instruction-level (i.e., SIMD) and task-level (i.e., multi-core) parallelism; and (ii) to generate efficient incremental code from the PAT, LightSaber uses a generalized aggregation graph (GAG), which encodes the low-level data dependencies required to produce aggregates over the stream. A GAG thus generalizes state-of-the-art approaches for incremental window aggregation and supports work-sharing between overlapping windows. LightSaber achieves up to an order of magnitude higher throughput compared to existing systems-on a 16-core server, it processes 470 million records/s with 132 ?s average latency.

Deep learning models are trained on servers with many GPUs, and training must scale with the number of GPUs. Systems such as TensorFlow and Caffe2 train models with parallel synchronous stochastic gradient descent: they process a batch of training data at a time, partitioned across GPUs, and average the resulting partial gradients to obtain an updated global model. To fully utilise all GPUs, systems must increase the batch size, which hinders statistical efficiency. Users tune hyper-parameters such as the learning rate to compensate for this, which is complex and model-specific. We describe Crossbow, a new single-server multi-GPU system for training deep learning models that enables users to freely choose their preferred batch size---however small---while scaling to multiple GPUs. Crossbow uses many parallel model replicas and avoids reduced statistical efficiency through a new synchronous training method. We introduce SMA, a synchronous variant of model averaging in which replicas independently explore the solution space with gradient descent, but adjust their search synchronously based on the trajectory of a globally-consistent average model. Crossbow achieves high hardware efficiency with small batch sizes by potentially training multiple model replicas per GPU, automatically tuning the number of replicas to maximise throughput. our experiments show that Crossbow improves the training time of deep learning models on an 8-GPU server by 1.3--4X compared to TensorFlow.

Modern servers have become heterogeneous, often combining multi-core CPUs with many-core GPGPUs. Such heterogeneous architectures have the potential to improve the performance of data-intensive stream processing applications, but they are not supported by current relational stream processing engines. For an engine to exploit a heterogeneous architecture, it must execute streaming SQL queries with sufficient data-parallelism to fully utilise all available heterogeneous processors, and decide how to use each in the most effective way. It must do this while respecting the semantics of streaming SQL queries, in particular with regard to window handling.

In this paper, we study the problem of multi-resource fairness in systems with multiple users. Each user requires to run one or more complex jobs that consist of multiple interconnected tasks. A job is considered finished when all its corresponding tasks have been executed in the system. Tasks can have different resource requirements. Because of special demands on particular hardware or software, tasks can have placement constraints limiting the type of machines they can run on. We develop User-Dependence Dominant Resource Fairness (UDRF), a generalized version of max-min fairness that combines graph theory and the notion of dominant resource shares to ensure multi- resource fairness between users with complex jobs. UDRF satisfies several desirable properties including strategy proofness, which ensures that users do not benefit from misreporting their true resource demands. We propose an offline algorithm that computes optimal UDRF allocation while the scheduling process can be to be decentralize across multiple schedulers. But optimality comes at a cost, especially for systems where schedulers need to make thousands of online scheduling decisions per second. Therefore, we develop a lightweight online algorithm that closely approximates UDRF. Large-scale simulations driven by Google cluster- usage traces show that UDRF achieves better resource utilization and throughput compared to the current state-of-the-art in multi-resource fair allocation.

Home wireless networks are difficult to manage and comprehend because of evolving locality, co-locality, connectivity and interaction. We define formal models of home wireless network infrastructure and policies and investigate how they can be used in a network management system designed to provide user-oriented support. We model spatial and temporal behaviour of network interactions and user-initiated network policies and define an online framework for generation of models from network and user-initiated events. The models are expressed in an extension to Milnerʼs bigraphical reactive systems. Analysis of the models is carried out in real-time by a bespoke bigraph reasoning system based on checking predicates, which is encoded as bigraph matching. Real-time model generation and analysis is implemented on the experimental Homework system router and trialled with synthetic and actual network data.

This paper presents a user driven redesign of the domestic network infrastructure that draws upon a series of ethnographic studies of home networks. We present an infrastructure based around a purpose built access point that has modified the handling of protocols and services to reflect the interactive needs of the home. The developed infrastructure offers a novel measurement framework that allows a broad range of infrastructure information to be easily captured and made available to interactive applications. This is complemented by a diverse set of novel interactive control mechanisms and interfaces for the underlying infrastructure. We also briefly reflect on the technical and user issues arising from deployments.

Home networks have evolved to become small-scale versions of enterprise networks. The tools for visualizing and managing such networks are primitive and continue to require networked systems expertise on the part of the home user. As a result, non-expert home users must manually manage non-obvious aspects of the network - e.g., MAC address filtering, network masks, and firewall rules, using these primitive tools. The Homework information plane architecture uses stream database concepts to generate derived events from streams of raw events. This supports a variety of visualization and monitoring techniques, and also enables construction of a closed-loop, policy-based management system. This paper describes the information plane architecture and its associated policy-based management infrastructure. Exemplar visualization and closed-loop management applications enabled by the resulting system (tuned to the skills of non-expert home users) are discussed.

The Homework project has examined redesign of existing home network infrastructures to better support the needs and requirements of actual home users. Integrating results from several ethnographic studies, we have designed and built a home networking platform providing detailed per-flow measurement and management capabilities supporting several novel management interfaces. This demo specifically shows these new visualization and control interfaces (1), and describes the broader benefits of taking an integrated view of the networking infrastructure, realised through our router's augmented measurement and control APIs (2).

Despite several advantages inherent in mobile-agent- based approaches to network management as compared to traditional SNMP-based approaches, industry is reluctant to adopt the mobile agent paradigm as a replacement for the existing manager-agent model; the management community requires an evolutionary, rather than a revolutionary, use of mobile agents. Furthermore, security for distributed management is a major concern; agent-based management systems inherit the security risks of mobile agents. We have developed a Java-based mobile agent infrastructure for network management that enables the safe integration of mobile agents with the SNMP protocol. The security of the system has been evaluated under agent to agent-platform and agent to agent attacks and has proved trustworthy in the performance of network management tasks.