ICWE 2021 Workshops

This paper proposes a composable “Just in Time Architecture” for Data Science (DS) Pipelines named JITA-4DS and associated resource management techniques for configuring disaggregated data centers (DCs). DCs under our approach are composable based on vertical integration of the application, middleware/operating system, and hardware layers customized dynamically to meet application Service Level Objectives (SLO - application-aware management). Thereby, pipelines utilize a set of flexible building blocks that can be dynamically and automatically assembled and re-assembled to meet the dynamic changes in the workload’s SLOs. To assess disaggregated DC’s, we study how to model and validate their performance in large-scale settings.

Deep learning-based visual analytic applications have drawn attention by suggesting fruitful combinations with Deep Neural Network (DNN) models and visual data sensors. Because of the high cost of DNN inference, most systems adopt offloading techniques utilizing a high-end cloud. However, tasks that require real-time streaming often suffer from the problem of an imbalanced pipeline due to the limited bandwidth and latency between camera sensors and the cloud. Several DNN slicing approaches show that effectively utilizing the edge computing paradigm effectively lowers the frame drop rate and overall latency, but recent research has primarily focused on building a general framework that only considers a few fixed settings. However, we observed that the optimal split strategy for DNN models can vary significantly based on application requirements. Hence, we focus on the characteristics and explainability of split points derived from various application goals. First, we propose a new simulation framework for flexible software-level configuration, including latency and bandwidth, using dockercompose, and we experiment on a 14-layered Convolutional Neural Network (CNN) model with diverse layer types. We report the results of the total process time and frame drop rate of 50 frames with three different configurations and further discuss recommendations for providing proper decision guidelines on split points, considering the target goals and properties of the CNN layers.

Internet of Things (IoT) integrates billions of smart devices and keeps growing. IoT technologies play a crucial role in smart applications that improve the quality of life. Likewise, the computational capacity of mobile devices has greatly increased, opening up new possibilities. In many cases, human interaction is necessary for IoT devices to perform properly. Users must configure more and more devices, investing time and effort. Artificial Intelligence (AI) techniques are currently used to predict user needs and behavior, trying to adapt devices to user preferences. However, achieving all-purpose models is a challenging task, aggravated by long training periods preventing personalized models in the early stages. This paper proposes a solution based on Federated Learning to predict behaviors in different environments and improve user’s coexistence with IoT devices, avoiding most manual interactions and making use of mobile devices capabilities. Federation allows new users’ predictions to be done using other users’ previous behaviors in similar environments. Also, it provides closer customization, immediate availability and avoids most manual device interactions.

For detecting anomalies which are unexpected behaviors in complex systems, deep learning-based anomaly detection algorithms for multivariate time series have gained a lot of attention recently. While many anomaly detection algorithms have been widely proposed, there has been no work on how to perform these detection algorithms for multivariate data streams with a stream processing framework. To address this issue, we present a real-time deep learning-based anomaly detection approach for multivariate data streams with Apache Flink. We train a LSTM encoder-decoder model to reconstruct a multivariate input sequence and develop a detection algorithm that uses reconstruction error between the input sequence and the reconstructed sequence. We show that our anomaly detection algorithm can provide promising performance on a real-world dataset. Then, we develop a Flink program by implementing three operators which process and transform multivariate data streams in a specific order. The Flink program outputs anomaly detection results in real time, making system experts can easily receive notices of critical issues and resolve the issues by appropriate actions to maintain the health of the systems.

Cloud computing emerges as a boon to business enterprises that offers increased productivity, economic efficiency and least operational and maintenance costs. The cost for the services offered is variable and dependent on market trends. The pricing model for cloud services are pay-per-use or subscription based as required circumstantially. With more and more provisions created for growing demands of resources, the pricing models are metering the requirements and are constantly moderating to provide optimal prices for services yet keeping best revenue schemes. To deliver business value to customers with QoS considerations within limits of infrastructure that they have, a price-wise categorization is required. In addition to this, the operational cost as well as infrastructure cost for cloud providers will differ insignificantly based-on whether one consumer or multiple consumers are serviced. The subscription model to service provisioning by the provider end is fueled up with discounts, price-cuts, offers and benefits to lure the customers thereby maximizing their resource utilization implicitly. Therefore, in order to grab the opportunity created by this competitive environment, we propose a dynamic pricing model for cloud brokers. In the existing cases of unexpected costs due to resource constraints going towards higher extremities, our model evaluates a price band for the customer’s transparency in cost and optimizes it. The benefits of the proposed pricing model is two fold. It provides the assurance about the maximum price that will be charged to the consumer while enjoying the existing benefits of dynamic pricing model; the price range calculation is an estimation done on the basis of resources requested, price history, current price, and risk premium are the decisive factors for our estimation rule. A fair proposition of resource allocation with maximal resource utilization, due returns for investments and an effective cost offering for the services is the aim of our work.

We consider an energy harvesting wireless sensor network (EH-WSN), where each sensor, equipped with the battery, senses its surrounding area. We first define the estimation error of the sensing data at a measuring point, which increases as the distance to the sensor increases and the age of information (AoI) of the data increases. The AoI is the elapsed time since the latest status is generated. We also define the network coverage, which is defined as the area having the estimation errors lower than a target value. As a performance metric, we use the \(\alpha \)-coverage probability, which is the probability that the network coverage is larger than a threshold \(\alpha \). Finally, in order to deal with dynamic and complex environments, we propose a reinforcement learning (RL) based algorithm which determines the activation of the sensors. In simulation results, we show the proposed algorithm achieves higher performance than baselines. In addition, we show the impact of the transmission power and the number of sensors on the \(\alpha \)-coverage probability.

Recently, graph neural networks (GNNs) have been improved under the influence of various deep learning techniques, such as attention, autoencoders, and recurrent networks. However, real-world graphs may have multiple types of vertices and edges, such as graphs of social networks, citation networks, and e-commerce data. In these cases, most GNNs that consider a homogeneous graph as input data are not suitable because they ignore the heterogeneity. Meta-path-based methods have been researched to capture both heterogeneity and structural information of heterogeneous graphs. As a meta-path is a type of graph pattern, we extend the use of meta-paths to exploit graph patterns. In this study, we propose TP-HAN, a heterogeneous graph attention network for exploiting triangle patterns. In the experiments using DBLP and IMDB, we show that TP-HAN outperforms the state-of-the-art heterogeneous graph attention network.

In this paper we revisit a taxonomy of client-side IoT software architectures that we presented a few years ago. We note that the emergence of inexpensive AI/ML hardware and new communication technologies are broadening the architectural options for IoT devices even further. These options can have a significant impact on the overall end-to-end architecture and topology of IoT systems, e.g., in determining how much computation can be performed on the edge of the network. We study the implications of the IoT device architecture choices in light of the new observations, as well as make some new predictions about future directions. Additionally, we make a case for isomorphic IoT systems in which development complexity is alleviated with consistent use of technologies across the entire stack, providing a seamless continuum from edge devices all the way to the cloud.