Predictable adaptation of network depths can be an effective way to control inference latency and meet the resource condition of various devices. However, previous adaptive depth networks do not provide general principles and a formal explanation on why and which layers can be skipped, and, hence, their approaches are hard to be generalized and require long and complex training steps. In this paper, we present a practical approach to adaptive depth networks that is applicable to various networks with minimal training effort. In our approach, every hierarchical residual stage is divided into two sub-paths, and they are trained to acquire different properties through a simple self-distillation strategy. While the first sub-path is essential for hierarchical feature learning, the second one is trained to refine the learned features and minimize performance degradation if it is skipped. Unlike prior adaptive networks, our approach does not train every target sub-network in an iterative manner. At test time, however, we can connect these sub-paths in a combinatorial manner to select sub-networks of various accuracy-efficiency trade-offs from a single network. We provide a formal rationale for why the proposed training method can reduce overall prediction errors while minimizing the impact of skipping sub-paths. We demonstrate the generality and effectiveness of our approach with convolutional neural networks and transformers.

Modern convolutional neural networks (CNNs) have massive identical convolution blocks, and, hence, recursive sharing of parameters across these blocks has been proposed to reduce the amount of parameters. However, naive sharing of parameters poses many challenges such as limited representational power and the vanishing/exploding gradients problem of recursively shared parameters. In this paper, we present a recursive convolution block design and training method, in which a recursively shareable part, or a filter basis, is separated and learned while effectively avoiding the vanishing/exploding gradients problem during training. We show that the unwieldy vanishing/exploding gradients problem can be controlled by enforcing the elements of the filter basis orthonormal, and empirically demonstrate that the proposed orthogonality regularization improves the flow of gradients during training. Experimental results on image classification and object detection show that our approach, unlike previous parameter-sharing approaches, does not trade performance to save parameters and consistently outperforms over parameterized counterpart networks. This superior performance demonstrates that the proposed recursive convolution block design and the orthogonality regularization not only prevent performance degradation, but also consistently improve the representation capability while a significant amount of parameters are recursively shared.

In many cyber-physical systems (CPS), one of the primary challenges is how to disseminate many sensor streams to collaborating entities in a timely, reliable, and scalable manner. However, in CPS, communication and related computation is often unstable and constitutes a serial bottleneck when multiple sensor streams need to be distributed. In this paper, we propose a novel publish-subscribe middleware architecture, called parallel data distribution service (PDDS), to address this problem. To bypass this serial bottleneck effectively, PDDS exploits state-space models for sensor streams both at a publisher and its subscribers. These models are used to estimate true states of monitored physical processes without actual communication. This approach not only reduces the communication overhead significantly, but also provides the tolerance against potential delays and instability in communication. Further, with the model-based methods, many sensor streams can be accessed in parallel since sensor streams can bypass a serial bottleneck, such as network stacks. At the middleware layer, PDDS processes the model-related computation, such as the state estimation, in parallel using modern embedded general-purpose computing on graphics processing units, enabling highly scalable and energy-efficient processing of massive sensor streams. The proposed approach is implemented in a data distribution service middleware using CUDA. Our evaluation results show that the proposed approach can achieve significant energy savings and much higher scalability compared to baseline approaches.

This brief presents a scalable, high-quality delay defect diagnosis method based on segment delay estimation. Several recent studies have assumed that accurate segment delay recovery leads to a better diagnosis method, and they predict segment delays using Gaussian prior distributions on them. We show that the assumption is not necessarily true and propose to rank segments by the probability of the occurrence for the estimated delays. When random localized defects are considered, prior distributions on segment delays can have a long tail, but all the previous studies fail to model this region properly. We propose to modify the standard deviations of the Gaussian priors depending on the defect size. Our experiment shows that one of our methods achieves ~14% first hit rank improvements and 100× speedup on average over a previous method based on linear programming.

There are growing demands to leverage network connectivity and interoperability of medical devices in order to improve patient safety and the effectiveness of medical services. However, if not properly designed, the integration of medical devices through networking could significantly increase the complexity of the system and make the system more vulnerable to potential errors, jeopardizing patient safety. The system must be designed and verified to guarantee the safety of patients and the effectiveness of medical services in the face of potential problems such as network failures. In this paper, we propose organ-centric hierarchical control architecture as a viable solution that reduces the complexity in system design and verification. In our approach, medical devices are grouped into clusters according to organ-specific human physiology. Each cluster captures common patterns arising out of medical device interactions and becomes a survivable semiautonomous unit during network failures. Further, safety verification and runtime enforcement can be modularized along organ-centric hierarchical control structure. We show the feasibility of the proposed approach under Simulink's model-based development framework. A simplified scenario for airway laser surgery is used as a case study.

Medical devices are increasingly capable of interacting with each other by leveraging network connectivity and interoperability, promising a great benefit for patient safety and effectiveness of medical services. However, ad-hoc integration of medical devices through networking can significantly increase the complexity of the system and make the system more vulnerable to potential errors and safety hazards. In this paper, we address this problem and introduce an organ-centric compositional development approach. In our approach, medical devices are composed into semi-autonomous clusters according to organ-specific physiology in a network-fail-safe manner. Each organ-centric cluster captures common device interaction patterns of sensing and control to support human physiology. The library of these formally verified organ-centric architectural patterns enables rapid and safe composition of supervisory controllers, which are specialized for specific medical scenarios. Using airway-laser surgery as a case study of practical importance, we demonstrate the feasibility of our approach under Simulink's model-driven development framework.

Medical Device Plug-and-Play (MDPnP) permits the mitigation of preventable medical errors. However, MDPnP results in a dynamic environment, which presents great challenges to traditional software architectures. This paper addresses these challenges by describing an abstract representation of dynamic clinical environment and a modular architecture is utilized to support safe system reconfiguration.

The Physically-Asynchronous Logically-Synchronous (PALS) system is a recently proposed architectural pattern for cyber-physical systems. It guarantees a logically synchronous design abstraction for real-time distributed computations. In this work, we develop a new middleware, called PALSware, to support an efficient and robust implementation of the PALS system and its extensions. PALSware guarantees consistency in distributed applications by eliminating any asynchronous interactions resulting from distributed clocks and node failures. We present a layered design for this middle-ware that is both reusable in different system architectures and can be extended with architecture-specific solutions for fault management. We demonstrate the middleware for an academic control testbed and show the consistency in a fault injection framework designed for this middleware.

Quality-aware realtime Embedded DataBase (QeDB) is a database for data-intensive real-time applications running on embedded devices. Currently, databases for embedded systems are best effort, providing no guarantees on their timeliness and data freshness. Existing real-time database (RTDB) technology cannot be applied to these embedded databases since it hypothesizes that the main memory of a system is large enough to hold the entire database. This, however, might not be true in data-intensive real-time applications. QeDB uses a novel feedback control scheme to support QoS in such embedded systems without requiring all data to reside in main memory. In particular, our approach is based on simultaneous control of both I/O and CPU resources to guarantee the desired timeliness. Unlike existing work on feedback control of RTDB performance, we implement and evaluate the proposed scheme on a modern embedded device. The experimental results show that our approach supports the desired timeliness of transactions while still maintaining high data freshness compared to baseline approaches.

One of the primary requirements in many cyber-physical systems (CPS) is that the sensor data derived from the physical world should be disseminated in a timely and reliable manner to all interested collaborative entities. However, providing reliable and timely data dissemination services is especially challenging for CPS since they often operate in highly unpredictable environments. Existing network middleware has limitations in providing such services. In this paper, we present a novel publish/subscribe-based middleware architecture called Real-time Data Distribution Service (RDDS). In particular, we focus on two mechanisms of RDDS that enable timely and reliable sensor data dissemination under highly unpredictable CPS environments. First, we discuss the semantics-aware communication mechanism of RDDS that not only reduces the computation and communication overhead, but also enables the subscribers to access data in a timely and reliable manner when the network is slow or unstable. Further, we extend the semantics-aware communication mechanism to achieve robustness against unpredictable workloads by integrating a control-theoretic feedback controller at the publishers and a queueing-theoretic predictor at the subscribers. This integrated control loop provides Quality-of-Service (QoS) guarantees by dynamically adjusting the accuracy of the sensor models. We demonstrate the viability of the proposed approach by implementing a prototype of RDDS. The evaluation results show that, compared to baseline approaches, RDDS achieves highly efficient and reliable sensor data dissemination as well as robustness against unpredictable workloads.

Achieving situation awareness is especially challenging for real-time data stream applications because they i) operate on continuous unbounded streams of data, and ii) have inherent real-time requirements. In this paper we show how formal data stream modeling and analysis can be used to better understand stream behavior, evaluate query costs, and improve application performance. We use MEDAL, a formal specification language based on Petri nets, to model the data stream queries and the Quality-of-Service (QoS) management mechanisms in a data stream system. MEDAL's ability to combine query logic and data admission control in one model allows us to design a single comprehensive model of the system. This model can be used to perform a large set of analyses to help improve the application's performance and QoS.

QeDB is a database for data-intensive real-time applications running on flash memory-based embedded systems.Currently, databases for embedded systems are best effort,providing no guarantees on its timeliness and data freshness. Moreover, the existing real-time database (RTDB) technology can not be applied to these embedded databases since they hypothesize that the main memory of a system is large enough to hold all database, which can not be true in data-intensive real-time applications. QeDB uses a novel feedback control scheme to support QoS in such embedded systems without requiring all data to reside in main memory.In particular, our approach is based on simultaneous control of both I/O and CPU resource to guarantee the desired timeliness. Unlike existing work on feedback control of RTDB performance, we actually implement and evaluate the proposed scheme in a modern embedded system.The experimental results show that our approach supports the desired timeliness of transactions while still maintaining high data freshness compared to baseline approaches.

It is a challenging task to provide timely access to global data from sensors in large-scale sensor network applications. Current data storage architectures for sensor networks have to make trade-offs between timeliness and scalability. PRIDE is a data abstraction layer for 2-tier sensor networks, which enables timely access to global data from the sensor tier to all participating nodes in the upper storage tier. The design of PRIDE is heavily influenced by collaborative real-time applications such as search-and-rescue tasks for high-rise building fires, in which multiple devices have to collect and manage data streams from massive sensors in cooperation. PRIDE achieves scalability, timeliness, and flexibility simultaneously for such applications by combining a model-driven full replication scheme and adaptive data quality control mechanism in the storage-tier. We show the viability of the proposed solution by implementing and evaluating it on a large-scale 2-tier sensor network testbed. The experiment results show that the model-driven replication provides the benefit of full replication in a scalable and controlled manner.

The demand for real-time data services in embedded systems is increasing. In these new computing platforms, using traditional buffer management schemes, whose goal is to minimize the number of I/O operations, is problematic since they do not consider the constraints of those platforms such as limited energy and distinctive underlying storage. In particular, due to asymmetric read/write characteristic of flash memory, minimum buffer misses neither coincide with minimum power consumption nor minimum I/O deadline miss ratio. In this paper we propose a power-aware buffer cache management scheme for real-time databases whose secondary storage is a flash memory. We focus on the problem of guaranteeing the performance goal in terms of both I/O power consumption and I/O deadline miss ratio. To address this problem, we propose logical partitioning of the global buffer pool into read and write buffer pools, and dynamic feedback control of read/write buffer pool sizes to satisfy both performance goals. We have shown through an extensive evaluation that our approach satisfies both performance goals in a variety of workloads and access patterns with considerably smaller size of buffer pools compared to baseline approaches.

Recently, cheap and large capacity non-volatile memory such as flash memory is rapidly replacing disks in embedded systems. While the access time of flash memory is highly predictable, deadline misses may occur if data objects in flash memory are not properly managed in real-time embedded databases. Buffer cache can be used to mitigate this problem. However, since the workload of a real-time database cannot be precisely predicted, it may not be feasible to provide enough buffer space to satisfy all timing constraints. Several deadline miss ratio management schemes have been proposed, but they do not consider I/O activities. In this paper, we present an I/O-aware deadline miss ratio management scheme in real-time embedded databases whose secondary storage is flash memory. We propose an adaptive I/O deadline assignment scheme, in which I/O deadlines are derived from up-to-date system status. We also present a deadline miss ratio management architecture where a control theory-based feedback control loop prevents resource overload both in I/O and CPU. A simulation study shows that our approach can effectively cope with both I/O and CPU overload to achieve the desired deadline miss ratio.

Environmental wireless sensor network (EWSN) systems are deployed in potentially harsh and remote environments where inevitable node and communication failures must be tolerated. LUSTER---Light Under Shrub Thicket for Environmental Research---is a system that meets the challenges of EWSNs using a hierarchical architecture that includes distributed reliable storage, delay-tolerant networking, and deployment time validation techniques.

The open grid services architecture (OGSA) addresses the need for standardization of diverse grid services by defining a set of core capabilities and behaviors needed by loosely coupled, service-oriented grid architectures. These OGSA standards and interfaces are based on ubiquitous, platform-neutral, technologies like SOAP, XML, and Web services. This paper briefly describes our OGSA proxy implementation that interconnects legacy Legion grids with grids that support the emerging OGSA specifications. Specifically we have constructed proxy services that support OGSA-ByteIO, WS-Directory, and WS-Naming