Razvan Racu

Applying sensitivity analysis in real-time distributed systems

During real-world design of embedded real-time systems, it cannot be expected that all performance data required for scheduling analysis is fully available up front. In such situations, sensitivity analysis is a promising approach to deal with uncertainties that result from incomplete specifications, early performance estimates, late feature requests, and so on. Sensitivity analysis allows the system designer to keep track of the flexibility of the system, and thus to quickly assess the impact of changes of individual hardware and software components on system performance. In this paper we integrate sensitivity analysis into our system-level performance analysis framework SymTA/S and show its benefits during the design of complex, networked multiprocessor embedded real-time systems.

Influence of different system abstractions on the performance analysis of distributed real-time systems

System level performance analysis plays a fundamental role in the design process of real-time embedded systems. Several different approaches have been presented so far to address the problem of accurate performance analysis of distributed embedded systems in early design stages. The existing formal analysis methods are based on essentially different concepts of abstraction. However, the influence of these different models on the accuracy of the system analysis is widely unknown, as a direct comparison of performance analysis methods has not been considered so far. We define a set of benchmarks aimed at the evaluation of performance analysis techniques for distributed systems. We apply different analysis methods to the benchmarks and compare the results obtained in terms of accuracy and analysis times, highlighting the specific effects of the various abstractions. We also point out several pitfalls for the analysis accuracy of single approaches and investigate the reasons for pessimistic performance predictions.

Scheduling analysis integration for heterogeneous multiprocessor SoC

Today, only very few techniques out of the host of work on formal performance and timing analysis have been adopted in MpSoC (multiprocessor system-on-chip) design. One of the key reasons is a mismatch between the scheduling models assumed in most formal approaches and the heterogeneous world of MpSoC scheduling techniques and communication patterns. This heterogeneity results from IP reuse and a plug-and-play design style, required to effectively reach the necessary design productivity. A second problem is the model complexity. While complex, specialized models can find their way into industry niches, their broad acceptance is extremely doubtful. In this paper, we review the existing scheduling analysis techniques with respect to these key requirements and derive a good compromise between model simplicity on the one hand, and applicability to MpSoC design on the other hand. The approach represents system-level scheduling analysis as a flow-analysis problem for event streams that can be configured to reuse the existing local scheduling analysis techniques. We define transformations between few key event stream models to meet the interfacing requirements of the compositional design style. An example demonstrates the application of the approach, as well as the worthiness of the results.

A formal approach to robustness maximization of complex heterogeneous embedded systems

Embedded system optimization typically considers objectives such as cost, timing, buffer sizes and power consumption. Robustness criteria, i.e. sensitivity of the system to variations of properties like execution and transmission delays, input data rates, CPU clock rates, etc., has found less attention despite its practical relevance. In this paper we introduce robustness metrics and propose an algorithm considering these metrics in design space exploration and system optimization. The algorithm can optimize for static and for dynamic robustness, the latter including system or designer reactions to property variations. We explain several applications ranging from platform optimization to critical component identification. By means of extensive experiments we show that design space exploration pursuing classical design goals does not necessarily yield robust systems, and that our method leads to systems with significantly higher design robustness.

Multi-dimensional Robustness Optimization in Heterogeneous Distributed Embedded Systems

Embedded system optimization typically considers objectives such as cost, timing, buffer sizes, and power consumption. Robustness criteria, i.e. sensitivity of the system to property variations like execution and transmission delays, input data rates, CPU clock rates, etc., has found less attention despite its practical relevance. In this paper we present an approach for optimizing multidimensional robustness criteria in complex distributed embedded systems. The key novelty of our approach is a scalable stochastic multi-dimensional sensitivity analysis technique approximating the sought-after sensitivity front from two sides, i.e. coming from the space of working and from the space of non-working system property combinations. We utilize the proposed stochastic sensitivity analysis to derive multi-dimensional robustness metrics, which are capable of bounding the robustness of given system configurations with little computational effort. The proposed metrics can significantly speed up multidimensional robustness optimization by quickly identifying promising system configurations, whose in-depth robustness evaluation can be performed subsequently to the optimization process

A formal approach to multi-dimensional sensitivity analysis of embedded real-time systems

System robustness is a major concern in the design of efficient and reliable state-of-the-art heterogenous embedded real-time systems. Due to complex component interactions, resource sharing and functional dependencies, one-dimensional sensitivity analysis cannot cover all effects that modifications of one system property may have on system performance. One reason is that the variation of one property can also affect the values of other system properties requiring new approaches to keep track of simultaneous parameter changes. In this paper we present a heuristic and a stochastic approach suited for the multi-dimensional sensitivity analysis of large heterogenous embedded systems with complex timing constraints

Sensitivity analysis of complex embedded real-time systems

Abstract The robustness of an architecture to changes is a major concern in the design of efficient and reliable state-of-the-art embedded real-time systems. Robustness is important during design process to identify if and in how far a system can accommodate later changes or updates, or whether it can be reused in a next generation product. In the product life-cycle, robustness helps the designer to perform changes as a result of product updates, integration of new components and subsystems, or modifications of the environment. In this paper we determine robustness as a performance reserve, the slack in performance before a system fails to meet timing requirements. This is measured as design sensitivity. Due to complex component interactions, resource sharing and functional dependencies, one-dimensional sensitivity analysis might not cover all effects that modifications of one system property may have on system performance. One reason is that the variation of one property can also affect the values of other system properties requiring new approaches to keep track of simultaneous parameter changes. In this paper we present a framework for one-dimensional and multi-dimensional sensitivity analysis of real-time systems. The framework is based on compositional analysis that is scalable to large systems. The one-dimensional sensitivity analysis combines a binary search technique with a set of formal equations derived from the real-time scheduling theory. The multi-dimensional sensitivity analysis engine consists of an exact algorithm that extends the one-dimensional approach, and a stochastic algorithm based on evolutionary search techniques.

Improved response time analysis of tasks scheduled under preemptive Round-Robin

Round-Robin scheduling is the most popular time triggered scheduling policy, and has been widely used in communication networks for the last decades. It is an efficient scheduling technique for integration of unrelated system parts, but the worst-case timing depends on the system properties in a very complex way. The existing works on response time analysis of task scheduled under Round-Robin determine very pessimistic response time bounds, without considering in detail the interactions between tasks. This may lead to a degradation of the efficiency of Round-Robin scheduling algorithm, and becomes a practical obstacle to its application in realtime systems. In this paper we present an approach to compute much tighter best-case and worst-case response time bounds of tasks scheduled under preemptive Round-Robin, including also the effects of the scheduling algorithm.

Scheduling Anomaly Detection and Optimization for Distributed Systems with Preemptive Task-Sets

Robustness and optimality are becoming the key principles in designing efficient and reliable state-of-the-art multi-processor real-time systems. However, due to complex inter-processor dependencies, the variation of local system parameters may have unpredictable system level impact including timing anomalies. In this context, the former heuristic optimization approaches used at resource level become less suitable for distributed systems with heterogeneous components and dynamic scheduling techniques. Techniques from exploration theory can better address the optimization problems but suffer from huge search spaces in general. In this paper, we present constructive methods for pointing out those system configurations that lead to anomalous behavior of the performance metrics. These are then used to guide the exploration process and reduce the search space, thereby increasing efficiency and making the approach applicable in practice. As a result, detailed information about anomalies can be quickly obtained and heavily exploited in system optimization, which we demonstrate using comprehensible examples.

Methods for power optimization in distributed embedded systems with real-time requirements

Dynamic voltagescaling and sleep state control have been shown to be extremely effective in reducing energy consumption in CMOS circuits. Though plenty of research papers have studied the application of these techniques in real-time embedded system design through intelligent task and/or voltage scheduling, most of these results are limited to relatively simple real-time application models. In this paper, a comprehensive real-time application model including periodic, sporadic and bursty tasks as well as distributed real-time constraints such as end-to-end delays is considered. Two methods are presented for reducing energy consumption while satisfying complex real-time constraints for this model. Experimental results show that the methods achieve significant energy savings without violating any deadlines.