Piotr Dziurzanski

A Survey and Comparative Study of Hard and Soft Real-Time Dynamic Resource Allocation Strategies for Multi-/Many-Core Systems

Multi-/many-core systems are envisioned to satisfy the ever-increasing performance requirements of complex applications in various domains such as embedded and high-performance computing. Such systems need to cater to increasingly dynamic workloads, requiring efficient dynamic resource allocation strategies to satisfy hard or soft real-time constraints. This article provides an extensive survey of hard and soft real-time dynamic resource allocation strategies proposed since the mid-1990s and highlights the emerging trends for multi-/many-core systems. The survey covers a taxonomy of the resource allocation strategies and considers their various optimization objectives, which have been used to provide comprehensive comparison. The strategies employ various principles, such as market and biological concepts, to perform the optimizations. The trend followed by the resource allocation strategies, open research challenges, and likely emerging research directions have also been provided.

Hard real-time guarantee of automotive applications during mode changes

This paper presents a resource allocation approach that benefits from modal nature of hard-real time systems under consideration. The modal nature determines the operational modes of the systems. Thanks to the modal nature of these systems, it is possible to decrease the number of active cores consuming high power in certain modes, leading to considerable energy savings while still not violating any of timing constraints. The proposed approach consists of both off-line and on-line steps. More computational intensive steps are performed off-line, whereas only detection of the current mode and mode switching are performed online. In the presented automotive use case, the number of required cores has been decreased up to 75% in a particular mode and relatively low amount of data is to be migrated during the mode change.

Market-inspired dynamic resource allocation in many-core high performance computing systems

Many-core systems are envisioned to fulfill the increased performance demands in several computing domains such as embedded and high performance computing (HPC). The HPC systems are often overloaded to execute a number of dynamically arriving jobs. In overload situations, market-inspired resource allocation heuristics have been found to provide better results in terms of overall profit (value) earned by completing the execution of a number of jobs when compared to various other heuristics. However, the conventional market-inspired heuristics lack the concept of holding low value executing jobs to free the occupied resources to be used by high value arrived jobs in order to maximize the overall profit. In this paper, we propose a market-inspired heuristic that accomplish the aforementioned concept and utilizes design-time profiling results of jobs to facilitate efficient allocation. Additionally, the remaining executions of the held jobs are performed on freed resources at later stages to make some profit out of them. The holding process identifies the appropriate jobs to be put on hold to free the resources and ensures that the loss incurred due to holding is lower than the profit achieved by high value arrived jobs by using the free resources. Experiments show that the proposed approach achieves 8% higher savings when compared to existing approaches, which can be a significant amount when dealing in the order of millions of dollars.

An interval algebra for multiprocessor resource allocation

This paper presents an interval algebra created specifically to evaluate timing properties of multiprocessor systems. It models the application load as intervals, and considers allocation and scheduling as algebraic operations over those intervals, aiming to analyse the impact of resource allocation decisions on application response times or schedulability. The theoretical background is introduced informally, followed by the description of a reference implementation of the interval algebra in C++, aiming to appeal to the design practitioner rather than the formalist. Examples of the usage of the proposed algebra are also provided, showing its applicability to the performance evaluation of industrial systems implemented over bus-based and Network-on-Chip multiprocessor platforms. A particular design flow is highlighted, where the interval algebra is used as a fitness function in a genetic algorithm tailored to optimise resource allocation in hard real-time multiprocessors.

Value and Energy Optimizing Dynamic Resource Allocation in Many-Core HPC Systems

The conventional approaches to reduce the energy consumption of high performance computing (HPC) data centers focus on consolidation and dynamic voltage and frequency scaling (DVFS). Most of these approaches consider independent tasks (or jobs) and do not jointly optimize for energy and value. In this paper, we propose DVFS-aware profiling and non-profiling based approaches that use design-time profiling results and perform all the computations at run-time, respectively. The profiling based approach is suitable for the scenarios when the jobs or their structure is known at design-time, otherwise, the non-profiling based approach is more suitable. Both the approaches consider jobs containing dependent tasks and exploit efficient allocation combined with identification of voltage/frequency levels of used system cores to jointly optimize value and energy. Experiments show that the proposed approaches reduce energy consumption by 15% when compared to existing approaches while achieving significant amount of value and reducing percentage of rejected jobs leading to zero value.

Design space exploration for complex automotive applications: an engine control system case study

With technological advances, significant changes are taking place in automotive domain. Modern automobile combines functionalities ranging from safety critical functions such as control systems for engine to navigation and infotainment. To meet the performances requirements of these systems, automotive industry is shifting to multi-core systems. This increases the design complexity. Efficient and fast design space exploration frameworks are required to deal with this design complexity. This paper presents a framework for exploring automotive application design on multi-core systems. It considers an automotive-specific application modeling language named Amalthea and a distributed-memory multi-core system architecture for execution. The effectiveness of our framework is shown on an engine control application.

Value and Energy Aware Adaptive Resource Allocation of Soft Real-Time Jobs on Many-Core HPC Data Centers

Modern high performance computing (HPC) data centers consume huge energy to operate them. Therefore, appropriate measures are required to reduce their energy consumption. Existing efforts for such measures focus on consolidation and dynamic voltage and frequency scaling (DVFS). However, most of them do not perform adaptive resource allocation for the executing dependent tasks (or jobs) in order to optimize both value and energy. The value is achieved by completing the execution of a job and it depends on the completion time. A high value is achieved if the job is completed before its deadline, otherwise a lower value. In this paper, we propose an adaptive resource allocation approach that uses design-time profiling results of jobs for efficient allocation and adaptation in order to optimize both value and energy while executing dependent tasks. The profiling results for each job are obtained by exploiting efficient allocation combined with identification of voltage/frequency levels of used system cores and used in adapting to different number of cores based on the monitored execution progress of the job and available cores. Experiments show that the proposed approach enhances the overall value by about 10% when compare to existing approaches while showing reduction in energy consumption and percentage of rejected jobs leading to zero value.

Feedback-Based Admission Control for Hard Real-Time Task Allocation Under Dynamic Workload on Many-Core Systems

In hard real-time systems, a computationally expensive schedulability analysis has to be performed for every task. Fulfilling this requirement is particularly tough when system workload and service capacity are not available a priori and thus the analysis has to be conducted at runtime. This paper presents an approach for applying control-theory-based admission control to predict the task schedulability so that the exact schedulability analysis is performed only to the tasks with positive prediction results. In case of a careful fine-tuning of parameters, the proposed approach can be successfully applied even to many-core embedded systems with hard real-time constraints and other time-critical systems. The provided experimental results demonstrate that, on average, only 62i?ź% of the schedulability tests have to be performed in comparison with the traditional, open-loop approach. The proposed approach is particularly beneficial for heavier workloads, where the number of executed tasks is almost unchanged in comparison with the traditional open-loop approach. By our approach, only 32i?ź% of exact schedulability tests have to be conducted. Moreover, for the analysed industrial workloads with dependent jobs, the proposed technique admitted and executed 11i?ź% more tasks while not violating any timing constraints.

Feedback-Based Admission Control for Firm Real-Time Task Allocation with Dynamic Voltage and Frequency Scaling

Feedback-based mechanisms can be employed to monitor the performance of Multiprocessor Systems-on-Chips (MPSoCs) and steer the task execution even if the exact knowledge of the workload is unknown a priori. In particular, traditional proportional-integral controllers can be used with firm real-time tasks to either admit them to the processing cores or reject in order not to violate the timeliness of the already admitted tasks. During periods with a lower computational power demand, dynamic voltage and frequency scaling (DVFS) can be used to reduce the dissipation of energy in the cores while still not violating the tasks’ time constraints. Depending on the workload pattern and weight, platform size and the granularity of DVFS, energy savings can reach even 60% at the cost of a slight performance degradation.

Feedback-based admission control for task allocation

This paper presents early exploration of the feedback admission control for high-performance computing clusters executing real-time tasks using controlled values to dynamic task allocation. A number of controller variants with different architectures and relying on various metrics have been proposed. Simulation models of both open- and closed-loop systems have been prepared and compared. The obtained experimental results show that the proposed approach leads to executing almost five times more tasks before their deadlines in case of periodic uniform workload, and about 16% more for bursty workloads with large computation time variance.