مدیریت منابع پویا بر اساس کنترل بازخورد در سیستم های زمان واقعی توزیع شده

The resource management in distributed real-time systems becomes increasingly unpredictable with the proliferation of data-driven applications. Therefore, it is inefficient to allocate the resources statically to handle a set of highly dynamic tasks whose resource requirements (e.g., execution time) are unknown a prior. In this paper, we build a distributed real-time system based on the control theory, focusing on the computational resource management. Specifically, this work makes three important contributions. First, it allows the designer to specify the desired temporal behavior of system adaptation, such as the speed of convergence. This is in contrast to previous literature, specifying only steady-state metrics, e.g. the deadline miss ratio. Second, unlike QoS optimization approaches, our solution meets performance guarantees with no accurate knowledge of task execution parameters – a key advantage in a poorly modeled environment. Last, in contrast to ad hoc algorithms based on intuition and testing, we rigorously prove that our approach not only has excellent steady state behavior, but also meets stability, overshoot, and settling time requirements.

Distributed real-time systems are widely used in highly dynamic environments where the resource requirements are open, fluctuating and not amenable to the traditional worst-case real-time analysis. For example, a web farm can be used to distribute time-sensitive contents such as movies and video clips. They need to handle a changing number of requests with significantly different resource requirements that are unknown beforehand. In a stock market, a system needs to actively push real-time stock updates at various interval to a group of users. The number of users served by a server can change quickly over time. Although these systems differ significantly in term of applications, they all operate in open environments where both workloads and available resources are difficult to predict. Monitoring and feedback control are needed to meet performance constraints. Several difficulties are observed in dynamic resource management in these systems. One main difficulty lies in their data-dependent resource requirements, which cannot be predicted without interpreting input data. For example, the execution time of an information server (a web or database server) heavily depends on the content of requests, such as the particular web page requested. A second major challenge is that these systems have highly uncertain arrival workloads; it is not clear how many users will request some resource in the web. A third challenge involves the complex interactions among many distributed sites, often across an environment with poor or unpredictable timing behavior. Consequently, developing certain types of future real-time systems will involve techniques for modeling the unpredictability of the environment, handling imprecise or incomplete knowledge, reacting to overload and unexpected failures (i.e., those not expressed by design-time failure hypotheses), and achieving the required performance levels and temporal behavior. We envision a trend in real-time computing to provide performance guarantees without the requirement of fine-grained task execution models, such as those depending on the precise estimation of individual task execution times. We shall see the emergence of coarse-grained models that describe the aggregate behavior of resource requirements. Coarse-grained models are easier to obtain and they need not be accurately computed. These models are more appropriate for dynamic resource management in the presence of uncertainties regarding load and resources. In this paper, we explore one such model based on difference equations. Unlike the more familiar queuing theory models of aggregate behavior, difference equation models do not make assumptions regarding the statistics of the load arrival process. Independent of the load assumptions, difference equation models are more suitable for systems where load statistics are difficult to obtain or where the load does not follow a distribution that is easy to handle analytically. The latter is the case, for example, with web traffic, which cannot be modeled by a Poisson distribution. Our solution has a basis in the theory and practice of feedback control scheduling. This is in contrast to the more common ad hoc resource management based on intuition and testing where it is very difficult to characterize the aggregate performance of the system and where major overloads and/or anomalous behavior can occur since these designs are not developed to avoid these problems.

To support data-driven applications with unpredictable and changing resource requirements, we develop here an effective computational resource management system, called DFCS, based on the feedback control. Different form other ad hoc approaches, DFCS has a basis in the theory. We have rigorously proven that our approach not only has excellent steady state behavior, but also meets stability, overshoot, and settling time requirements. We have demonstrated that DFCS is a better option for distributed resource management, than QoS (Adbelzaher et al., 2000) and DQM (Brandt et al., 1998).