Adapting and evaluating distributed real-time and embedded systems in dynamic environments

Abstract

Quality of Service (QoS)-enabled publish/subscribe (pub/-sub) middleware provides much needed infrastructure for data dissemination in distributed real-time and embedded (DRE) systems. It is hard, however, to quantify the performance of mechanisms that support multiple interrelated QoS concerns, e.g., reliability, latency, and jitter. Moreover, once an appropriate mechanism is selected, it is hard to maintain QoS properties as the operating environment fluctuates since the chosen mechanism might no longer provide the needed QoS. For DRE systems operating in such environments, adjustments to mechanisms supporting QoS must be both timely and resilient to unforeseen environments. This paper describes our work to (1) define composite metrics to evaluate multiple interrelated QoS concerns and (2) analyze various adjustment techniques ( i.e., policy-based approaches, machine learning techniques) used for the QoS mechanisms of a DRE system in a dynamic environment. Our results show that (1) composite metrics quantify the support that mechanisms provide for multiple QoS concerns to ease mechanism evaluation and creation of related composite metrics and (2) neural network machine learning techniques provide the constant-time complexity needed for DRE pub/-sub systems to determine adjustments and the robustness to handle unknown environments.

Adapting Distributed Real-Time and Embedded Pub/Sub Middleware for Cloud Computing Environments

“…ADAMANT extends our prior work [10,11] by empirically evaluating (1) the QoS delivered by DDS pub/sub middleware with respect to differences in computing and networking resources provided by cloud environments and (2) the accuracy and timeliness of ANN-based machine learning tools in determining appropriate middleware configurations. Figure 3 shows how ADAMANT works in a cloud environment (e.g., the ad-hoc SAR datacenter) to deploy cloud resources.…”

“…We selected ANN technology [11] due to its (1) fast and predictable performance, (2) accuracy for environments known a priori (i.e., used for ANN training) and unknown until runtime (i.e., not used for ANN training), and (3) low accidental development complexity. In particular, we chose the Fast Artificial Neural Network (FANN)(leenissen.dk/fann) implementation due to its configurability, documentation, ease of use, and open-source code.…”

“…Our prior work [10,11] on QoS-enabled pub/sub middleware performance for noncloud environments indicated that some transport protocols provide better reliability (as measured by the number of network packets received divided by the number sent) and latency for certain environments while other protocols are better for other environments. We therefore developed several composite QoS metrics to evaluate multiple QoS aspects simultaneously, thereby providing a uniform and objective evaluation of ADAMANT in cloud computing environments.…”

“…Our prior work [10,11] developed composite metrics to evaluate pub/sub middleware with various ANT-based transport protocols based on differences in application parameters (e.g., number of data receivers and data sending rate). We also evaluated multiple approaches for adapting to application parameter changes in a dedicated (i.e., non-cloud) operating environment without regard to changes in computing or networking resources.…”

Evaluating Supervised Machine Learning for Adapting Enterprise DRE Systems

“…These techniques are not well-suited for enterprise DRE systems, however, since they do not provide bounded times when determining adjustments [4]. Some techniques, such as reinforcement learning [6], explore the solution space until an appropriate solution is found, regardless of the elapsed time.…”