ABSTRACT. Coarse Grained Reconfigurable Arrays (CGRA) have become a popular technology to realize compute accelerators. CGRAs can be found in High-Performance systems and also in embedded systems. In order to provide the highest speedup, they need to support conditional statements and nested loops. This requires a management of conditions within the CGRA. This management can be done in different ways. In this contribution, we compare two such concepts and evaluate the impact that these concepts have on the achievable clock frequency, the required resources and the change of schedules. It turns out, that with our new condition management and the accompanying advanced schedule, we can save more than 17% of runtime.

ABSTRACT. Homogeneous computing systems are reaching their limits with the growing demands of current applications. Accelerating compute-intensive applications ensures manageable computing times and boosts energy efficiency, which is an important lever as part of ongoing efforts to tackle global climate change. Field Programmable Gate Array (FPGA) accelerators are well-known for increasing throughput and, in particular, energy efficiency for many applications. FPGA accelerators connected directly to the data center high-speed network are ideal for integration into a heterogeneous data center, avoiding the energy and resource overhead of a carrier system. The standalone Network-attached Accelerators (NAAs) further benefits from low latency and predictable line-rate network throughput, as well as an interoperable communications interface. For selected use cases, we compare a heterogeneous computing cluster extended by NAAs with a homogeneous CPU-based cluster not only in terms of computing performance and energy efficiency, but also considering resource efficiency. For this purpose, we perform a Life Cycle Assessment (LCA) for both systems based on the Key Performance Indicators for Data Center Efficiency (KPI4DCE) indicator set, which takes into account the manufacturing phase in addition to the usage phase. The KPI4DCE tool has been extended to include modeling of NAAs. This allows us to show that NAAs are not only more energy-efficient, but also more resource-efficient for the selected applications, leading to a strong improvement of the environmental impact of the manufacturing phase.

ABSTRACT. With the growing popularity of Processing-In-Memory (PIM) technology, many sectors of the industry are willing to take advantage of this new technology. However, the state-of-the-art applications are not optimized to fully utilize the PIM capabilities. In this paper, an in-memory database is analyzed and its functions whose executions cause the majority of CPU clock cycles are identified. Factors such as running time and cache locality are studied and processes causing long running times are accelerated with the PIM technology. The results show that by utilizing the proposed optimization methods, there is an overall speedup of 110.94\% in the selected functionalities in the database management system. Furthermore, a deep analysis of the results is provided, summarizing key observations and programming recommendations for the in-memory database developers, and providing guidelines on where to take advantage of this new memory technology, and where to avoid it.

ABSTRACT. The increasing demand for real-time data transmission for the remote-controlled units and the complexity of 5G networks pose significant challenges to achieving optimal performance in device-based scenarios, when the 5G network cannot be controlled by its users. This paper proposes a model-free Deep Reinforcement Learning approach for this task. The model learns an optimal policy for maximizing the data transmission rate while minimizing the latency and packet loss. Such an approach aims to investigate the applicability of the environment-agnostic agents driven purely by the transmission statistics of the acknowledged packets. The evaluation is done with the help of a 5G simulation based on the OMNeT++ network simulator and the obtained results are compared to a classic throughput-based adaptive bitrate streaming approach. Multiple questions and challenges that arose on the way to the final model and evaluation procedure are highlighted in detail. The resulting findings demonstrate the effectiveness of Deep Reinforcement Learning for optimizing real-time data transmission in 5G networks in an online manner.

ABSTRACT. Marine Autonomous Surface Ships (MASS) are gaining interest worldwide with the potential to reshape mobility and freight transport at sea. Collision avoidance and path planning are central components of the intelligence of a MASS. Deep Reinforcement Learning (DRL) techniques often learn these capabilities in a simulated environment. This article investigates how to learn collision avoidance and path planning solely from observational data, reducing the need for a simulator for training. A state-action dataset of ship trajectories is constructed from recorded Automatic Identification System (AIS) messages. Using this data, we investigate an application of the Prediction and Policy Learning Under Uncertainty (PPUU) technique. This includes training an action-conditional forward model and learning a policy network by unrolling future states and back-propagating errors from a self-defined cost function. To evaluate the learned policy, FerryGym, a Gymnasium environment is developed for evaluating the policy network from observational data.

ABSTRACT. Organic Computing (OC) is a well-known research field aiming to build dependable embedded systems. OC systems often employ self-X properties such as self-configuration, self-healing, etc. These properties are inherent to several biological systems such as the human body and offer a blueprint for technical systems.

The Artificial DNA (ADNA) system was developed in the scope of the OC research. Its basic idea is to build a dependable embedded system from a textual description (the artificial DNA -- as a technical counterpart to the DNA in biological cells).

Our contribution in this paper is to use the ADNA system to realize a highly dependable drone swarm providing self-X properties. We describe details of our drone demonstrator which we built for this purpose. In addition, we describe the extensions on the ADNA system to realize functions such as path planning and swarm control. The evaluation considers time delays in the WiFi connection between drones and Ground Control Stations (GCSs) and demonstrates that the real-time requirements of the ADNA system mostly hold despite the delays.