1. Introduction2. Motivation3. Project summary4. Project details4.1. Architecture 4.2. New challenges4.3. Contributions4.4. Timeline5. ConclusionsReferencesSolving Energy-Latency Dilemma: Task Allocation for Parallel Applications in Heterogeneous Embedded Systems 1. Introduction Parallel applications with energy and low-latency constraints are emerging in various networked embedded systems like digital signal processing, vehicle tracking, and infrastructure monitoring. However, conventional energy-driven task allocation schemes for a cluster of embedded nodes only concentrated on energy-saving when making allocation decisions. Consequently, the length of the schedules could be very long, which is unfavorable or in some situations even not tolerated. In this project, we address the issue of allocating a group of parallel tasks on a heterogeneous embedded system with an objective of energy-saving and short-latency. A novel task allocation strategy, or BEATA (Balanced Energy-Aware Task Allocation), is developed to find an optimal allocation that minimizes overall energy consumption while confining the length of schedule to an ideal range. The rest of the proposal is organized as follows. In the next section we describe the motivation of this research. In Section 3, we propose the project summary. Project details are presented in Section 4. Section 5 concludes the proposal with summary and future directions. 2. Motivation Extensive researches have been conducted to reduce overall energy consumption for a variety of embedded systems using diverse techniques [1][2][3]. In particular, most of recent researches in energy-saving for embedded systems share two common features (1) applications considered are real-time in nature with hard deadlines; and (2) energy-saving is achieved by employing DVS (Dynamic Voltage Scaling). Our work is fundamentally different from the above approaches as we focus on reducing both energy consumption and response time for soft real-time parallel applications running on heterogeneous embedded systems with no DVS available. In a heterogeneous embedded system, different processing nodes have distinct fixed energy consumption rates. Similarly, different communication channels also have various energy assumption rates. The goal of this work is to develop a task allocation strategy that not only conserves energy but also generates a short schedule, which is favorable or even necessary in some scenarios. For example, in a soft real-time embedded system such as a cellular phone [2], it must be able to encode outgoing voice and decode incoming signal during a conversation in a timely manner. Occasional glitches in conversations due to tardy response are not desired. When the response time becomes longer frequent glitches could happen, which are not tolerated at all. Energy-saving and low-latency, however, are two conflicting objectives in the context of allocating a parallel application represented by a task graph onto a set of connected heterogeneous processing nodes in an embedded system. The dilemma arises from a multidimensional heterogeneity bearing by a heterogeneous embedded system. Specifically speaking, a processing node that provides a task with earliest finish time may not be an ideal candidate in terms of energy-saving. This is because the execution time of a task allocated on an embedded node is irrelative to the energy consumption rate offered by the node. Moreover, the computational energy consumption of a task allocated on a node is a product of energy consumption rate of the node and execution time of the task. The motivation of this work is to solve the energy-latency dilemma existed in networked heterogeneous embedded system where both energy-saving and low-latency need to be achieved. In this project, we address the issue by minimizing energy consumption while confining schedule lengths. To this end, we devised an energy-adaptive window to control the trade-off between energy consumption and response time. 3. Project summary In this project, we address the issue of allocating tasks of parallel applications in heterogeneous embedded systems with an objective of energy-saving and latency-reducing. BEATA (Balanced Energy-Aware Task Allocation), a task allocation scheme considering both energy consumption and schedule length, is developed to solve the energy-latency dilemma. To facilitate the presentation of BEATA, we will also propose mathematical models to describe a system framework, parallel applications with precedence constraints, and energy consumption model. Extensive simulations using a real world application as well as synthetic benchmarks will be conducted to compare the performance of existing approaches with that of the BEATA scheme. The experimental results will show that BEATA significantly improves the performance in terms of 1t1 ECN1=12.6 energy dissipation and makespan time over two baseline allocation schemes. 4. Project details 4.1. Architecture A networked embedded system in the most general form consists of a set, e.g., P = {p1, p2, ..., pm}, of heterogeneous embedded computing nodes (hereinafter referred to as nodes or embedded nodes) connected by a single-hop wired or wireless network. The network embedded system can be represented as a graph of nodes along with their point-to-point links. In the system, an embedded node is modelled as a vertex. There exists a weighted edge between two vertices if they can communicate with each other. Each node in the system has an energy consumption rate measured by Joule per unit time. With respect to energy conservation, each network link is characterized by its energy consumption rate that heavily relies on the link’s transmission rate, which is modelled by weight wij of the edge between node pi and pj. An allocation matrix X is an n×m binary matrix used to reflect a mapping of n tasks to m embedded nodes. Element xij in X is “1” if task ti is assigned to node pj and is “0”, otherwise. Heterogeneity investigated in this study embrace multiple meanings. First, execution times of a task on different embedded nodes may various, since the nodes may have different processing capabilities. Second, a node offering task ti a shorter execution time does not necessarily provide another task tj with a shortened execution time, because different nodes may have distinct processor architectures. This implies that different nodes in a system are suitable for different kinds of