CSR: Small: Timely Power-Aware Data Management in Embedded Systems

点击翻译按钮获取中文摘要

In emerging embedded and cyber physical applications like smart cars, micro grids, and homeland security, data amounts are increasing fast. Database support is essential in data-intensive embedded applications, because developing them without database support is hard and error-prone. Ideally, a database system needs to process real-time data service requests, such as driving route recommendations and electricity demand/supply analysis, in a timely manner using fresh data representing the current real-world status, e.g., the present traffic or electric grid status. It is also important the database should consume minimal power considering stringent power constraints in embedded systems. Achieving this vision is challenging for several reasons, including dynamic workloads varying in time depending on the current real world status, severe data/resource contention, and computational costs for updating temporal data. Moreover, the timeliness, data freshness, and power efficiency may compete with each other. If higher priority is given to user queries, their timeliness can be improved at the cost of the decreased data freshness or vice-versa. Simply consuming more power to support the desired timeliness and data freshness is not desirable. Unfortunately, state-of-the-art database systems may fall short of power-aware real-time data services. Non-real-time databases unaware of timing and data freshness requirements may perform poorly in these applications. Most existing real-time and embedded databases can provide no guarantee on the desired timeliness and data freshness. Neither are they power-aware. This can be a serious problem, since the unbounded tardiness or data staleness may result in a traffic jam, power outage, or homeland security problem. Although research on energy-efficient databases mainly in data center contexts has recently begun, real-time deadlines and data freshness requirements are not considered. Despite the importance, related work on power management in real-time databases, chiefly targeting embedded applications, is surprisingly scarce. Due to the power ignorance, deploying real-time databases may become increasingly difficult and costly. To bridge the gap, in-depth research will be performed in this work to investigate effective fundamental approaches for power-aware real-time data services in emerging embedded and cyber physical systems with significant broader impacts. This project will investigate novel methods to support the desired timeliness and data freshness even in the presence of dynamic workloads, while substantially decreasing the power consumption in real-time embedded databases (RTEDBs). This will involve exploring 1) application of adaptive control theoretic techniques to gracefully adapt the RTEDB system behavior, if necessary, to support the desired timeliness and data freshness for less power in the closed-loop system; 2) feedforward methods to avoid data/resource contention incurring performance and power penalties by considering RTEDB semantics; 3) a multicore RTEDB system architecture, in which the cores support the desired timeliness and data freshness, decreasing the power usage both autonomously and collaboratively; and 4) extensive simulations and experiments in a testbed will be performed using real-world data traces to measure the timeliness, data freshness, and power and energy consumption.