关键词： 能耗优化   MPSoC   能耗模型   高速缓存   通信优化  

摘要： 近年来仅仅依赖于制造工艺改进的集成电路已不能满足市场的需求,设计者们以复杂度和集成度更高的片上系统(SOC)来应对。多处理器片上系统(MPSoC)己成为嵌入式系统的主流芯片解决方案,其极高的复杂度和集成度带来严重的能耗问题,给设计者带来重大挑战。在SoC的设计流程中,既要需要对各系统模块进行能耗评估,并根据评估结果调整架构；同时还要基于各种高能效方法设计系统策略,设计相应的软硬件来使得系统运行时达到最优的能效。很多研究都集中在低功耗策略上,比如动态电压频率调整(DVFS)和时钟门控(Clock Gate)技术的研究。但是,低功耗并不意味着高能效,能效体现了能量的使用效率,体现了系统在单位能量下所能做的工作量。单纯的低功耗技术往往导致系统运行时间的延长,在延时超过一定程度则反而降低了系统能效。为此,在考虑系统能效的时候需要将应用的执行时间考虑进去,对系统中的各个模块的能耗做定量的分析。其次,这些低功耗技术都基于较低层次的电路,比如寄存器转换级(RTL)和门级(Gate),属于SoC设计的后期阶段。随着MPSoC的设计复杂度提高,其开发周期加长,在RTL级和门级的能耗评估和设计更加耗时。由于这两方面的原因,如果SoC设计后期阶段的能耗评估指标不满足规格要求,那么系统架构调整将十分的棘手,因此需要研究更早期的能耗评估与能耗优化技术。为此,本文深入研究MPSoC系统的高能效技术,结合MPSoC中各个子系统自身特点,将复杂问题局部化,根据各子系统的特点来制定相应的高能效方案。本文的研究包含以下三个方面：1)处理器子系统能耗评估和优化技术研究。为了早期能对处理器子系统在执行系统应用时的能耗进行评估,本文首先提出一种改进的指令层次的处理器能耗模型,并给出综合使用剖析和标注技术的能耗评估方法,该方法的MPSoC建模在虚拟架构(VA)层和传输精确(TA)层进行,兼顾评估的快速性和准确性。使用通用的代码分析工具GNU gcov来剖析VA层仿真时的统计数据,利用VA层仿真的速度,快速得到分析系统应用能耗的所需信息。然后,将VA层得到的信息标注到TA层模型,利用TA层模型的更高精确度来进一步评估处理器子系统能耗。最后,基于能耗分析结果,对以H.264/AVC解码应用驱动的MPSoC处理器子系统进行了优化,将运算复杂度过高的逻辑硬化成硬件加速器,在加速器的硬件实现过程中充分应用能耗优化架构和技术。2)存储子系统能耗优化技术研究。片上高速缓存(Cache)在存储子系统中扮演着重要角色,贡献了MPSoC的大部分能耗。本文提出一种对访问能自适应和线程感知相结合的Cache分割策略。该策略中,针对多核处理器的数据类型,本文提出将Cache中的数据区分为独享和共享,并根据数据类型区分访问类型。然后提出一种支持Cache分割的软硬件架构,并综合考虑访问类型、应用程序量和数据量,通过分割Cache给系统应用分配和使用适量的存储资源。进一步地,本文将应用线程数纳入分析模型,使用整数线性规划(ILP)的方法获取最优能耗结果和各处理器核心(Core)对应的Cache分配。3)通信子系统能耗优化技术研究。为了避免任务映射、线程划分及调度对通信子系统能耗的影响,本文首先提出一种综合使用消息聚合和通信流水线技术的策略。该策略中,本文采用消息聚合技术来减少通信次数,从而减少因启动通信而产生的能耗。本文还采用通信流水线技术来隐蔽通信开销,从而减少因等待通信而产生的能耗。进一步的,本文对消息聚合和通信流水线技术进行了量化分析,并采用基于ILP的方法确定通信流水线技术和消息聚合技术的使用,利用ILP方法的特性进行通信优化并获得最优能耗结果。其次,本文针对多核之间的通信协议,提出一种对称多核的高效一致性侦听过滤技术和装置,该技术通过分析和过滤冗余的Cache一致性操作,从而减少通信系统的拥塞和冗余通信,进一步可以减少通信子系统的能耗。

关键词： 硬实时系统   全局调度   抢占阈值   限制抢占   容错调度   主副版本  

摘要： 硬实时系统要求系统中的每个实时任务都能够在截止期之前正确响应,否则会带来难以承受的严重后果。硬实时调度算法构造确定性的任务运行秩序,并从理论上证明该运行秩序的可行性,确保任务不会错失截止期。随着硬实时系统对计算能力需求的不断增长,越来越多的复杂硬实时系统开始采用多处理器芯片。和单处理器实时系统相比,多处理器系统由于任务的并行运行,在调度方面的分析、设计难度较大。因此,高效率的多处理器实时调度算法是硬实时系统研究的核心内容之一抢占在提高调度能力和灵活性的同时会带来额外的系统开销。特别在多处理器平台上,抢占带来的作业迁移会造成不可忽视的性能下降。以减少不必要抢占为目标,提出多处理器实时系统的全局抢占阈值调度算法GPTS。GPTS是处于抢占调度和不可抢占调度之间的一种混合调度算法。该算法在作业开始运行时提升其优先级,从而降低了该作业被抢占的可能性,在保持系统可调度的基础上,可以有效减少系统运行过程中的抢占次数。限制不必要的抢占不仅可以减少资源损耗,还能够提高调度能力。将任务整体提升优先级会给高优先级任务带来过高的额外干扰,影响高优先级任务的实时性。以提高调度能力为目标,提出多处理器硬实时系统的限制抢占调度算法FPLPS.该算法在任务的末尾设置一段有较高优先级的区域,减少任务在该区域内运行时被抢占的次数,同时避免给资源压力很大的高优先级任务带来过多的额外干扰,从而提高了实时任务集的可调度性。硬实时系统在运行过程中,不可避免的会出现故障,在出现故障时依然可以正确运行是硬实时系统必须满足的要求。主副版本机制是硬实时调度中广泛采用的容错方法。在主副版本机制的容错调度算法中,主版本出现故障之后对应的副版本才会进入就绪状态,可供副版本运行的时间区间短,容易错失截止期。针对副版本实时性较差的问题,提出副版本不可抢占的全局容错调度算法FTGS-NPB。该算法通过赋予副版本最高优先级,使副版本在运行过程中不受干扰并在最短时间内响应,改善了副版本的实时性,从而减少了实现容错所需的额外处理器资源。副版本根据进入就绪状态的时间分为主动副版本和被动副版本。主动副版本在无故障时会消耗大量的额外资源,被动副版本不消耗额外资源,但是实时性较差,并且不适用于主副版本总使用率超过1的任务。同时考虑两种副版本的优缺点,提出副版本延迟的全局容错调度算法FTGS-BD。该算法根据资源使用情况将副版本的就绪时间尽量推迟,同时在不需要副版本响应时及时回收分配给副版本的资源,在保证副版本能够在对应主版本出现故障的情况下正确响应的同时,减少了副版本在主版本无故障时浪费的处理器资源,从而以较少的代价实现容错调度。同时,FTGS-BD可以应用于存在主副版本总使用率超过1的任务的实时任务集。

关键词： 动态部分可重构   离线可重构   可重构计算   智能计算   按需计算  

摘要： 可重构系统可以根据应用程序运行时的需要动态地选择不同配置硬件加速方案。动态可重构多处理器片上系统是一种面向多种应用/标准之间资源共享、减少能源损耗和提升系统性能(专用DSP或者定制协处理器)的有效方法。提出一种根据应用不同需求进行动态软硬件重构的方法。

关键词： Competitive ratio   Improved rejection penalty   Makespan   Off-line scheduling   On-line scheduling   Rejection penalty  

摘要： This paper deals with multiprocessor scheduling with rejection technique where each job is provided with processing time and a given penalty cost. If the job satisfies the acceptance condition, it will schedule in the least loaded identical parallel machine else job is rejected. In this way its penalty cost is calculated. Our objective is to minimize the makespan of the scheduled job and to minimize the sum of the penalties of rejected jobs. We have merged 'CHOOSE 'and 'REJECTION PENALTY' algorithm to reduce the sum of penalties cost and makespan. Our proposed 'Improved Reject penalty algorithm' reduce competitive ratio, which in turn enhances the efficiency of the on-line algorithm. By applying our new on-line technique, we got the lower bound of our algorithm is is 1.286 which is far better from the existing algorithms whose competitive ratio is at 1.819. In our approach we have consider non-preemption scheduling technique. Copyright © 2015 Institute of Advanced Engineering and Science. All rights reserved.

关键词： 遥感图像   图像加速器   流多处理器   结构设计   基本指令系统  

摘要： GPU(Graphic Processor Unit)是一种拥有几何编程能力和像素处理能力的图像芯片。流多处理器是GPU中最基本的运算单元，是GPU的核心部分，流多处理器的设计关系到GPU的整体计算性能。在航空航天应用中，遥感图像处理具有计算量大、数据量大的问题，在FPGA中设计高速低功耗的流多处理器组成高性能的GPU软核来加速遥感图像的处理具有重要意义和潜在的应用价值。本文针对GPU中流多处理器的设计技术开展研究，主要内容如下:　　(1)对GPU体系结构的发展历史和国内外研究现状进行了综述，介绍了GPU体系结构特点、开发平台及主要应用领域。结合遥感图像并行化处理的迫切需求，论述了研究GPU中流多处理器体系结构的重要意义。深入分析了NVIDIA G80/GT200和AMD R600的GPU体系结构，并对NVIDIA PTX指令集和AMD R600指令集进行了总结和分析。　　(2)在NVIDIAG80系列GPU中流多处理器架构的基础上，设计了一种流多处理器的基本结构，包括流水线的设计和存储系统的设计。对流多处理器执行指令时遇到的warp分歧问题和线程切换问题进行了深入的研究及实现。　　(3)设计了一种流多处理器的基本指令系统。给出了指令系统的设计文档，介绍了指令系统中的指令编码格式、寻址方式、寄存器的定义、程序状态字及条件码的定义。对各类指令的格式、编码和功能进行了详细地解释。　　(4)对流多处理器进行了仿真验证，并对仿真波形进行了详细地解释与分析。仿真波形结果表明，流多处理器运行正常，设计方法正确可行，取得了预期的结果，成功实现了各种指令的功能，解决了warp分歧和线程切换等关键问题。

关键词： Graphics processing unit  

摘要： In this paper, a Parallel Multiprocessor (PM) based on SIMT (Single Instruction and Multiple Threads) architecture is proposed. With co-issue architecture and multi-dimensional parallelism implemented in high-effective PM, Embedded Graphics Processing Unit (EGPU) provides great performance for various situations, such as general purpose computing, 3D scene rendering, and graphics processing. Application programs are departed into separated threads. Allocated by Thread Processing Unit (TPU), separated threads can be executed in parallel. Parallelism in different hierarchy and dimension are implemented by Multi-Dimensional Parallelism Processor (MD-PP), which has made a proper trade-off between performance and cost. Additionally, PM improves the hardware occupancy with its co-issue architecture and internal bus accessing mechanism to meet the demand of processing capability. Its unified shading architecture also helps to hide processing latency. PM can execute 4 basic operations in the best case and 2 in the worst case within each clock cycle. With 32nm process technology and 200MHz clock frequency, PM’s area is about 5104494um2, power consumption is about 101.838mW, and it can process nearly 28M vertices or fragments in average. Experimental results show that the MD-PP based PM can process data with high performance and get a balance between efficiency and hardware consumption simultaneously. © 2015 SERSC.

关键词： 棒材轧制   穿水冷却   分布式多处理器   闭环控制   上位机  

摘要： 棒材的生产正朝着连续化、高速化方向发展,这就决定了棒材在轧制后要能够实现快速冷却,以提高生产效率。目前应用穿水冷却技术实现棒材轧制过程中的冷却已得到业内的认可。不过传统的穿水冷却控制系统自动化程度较低,特别在中小型钢铁企业中,控制方式一般为单机开环控制,只能实现简单的单任务顺序控制,某些工序还需要人工操作,很难保证较高的控制精度和效率,更难满足棒材冷却控制的快速性和稳定性。本文针对棒材穿水冷却工艺的原理、特点及工艺流程展开研究,分析穿水冷却过程中的核心控制因素,即冷却水的压力、温度及流量和棒材轧件在穿水前后的温度。结合目前分布式多处理器控制技术的发展状况,提出了棒材穿水冷却的分布式多处理器闭环控制方案。这种控制方式结合了分布式控制和多处理器的优点,将分布式多处理器的思想应用于棒材穿水冷却系统中,设计了棒材穿水冷却控制系统的体系结构。通过对棒材穿水冷却控制环节的研究,对现场软硬件进行选型设计;并通过西门子PLC编程软件STEP 7完成了棒材穿水冷却闭环控制系统的程序设计;通过西门子组态软件WINCC设计了监控系统。棒材穿水冷却分布式多处理器闭环控制系统可以根据穿水后轧钢的温度来对系统进行闭环控制,实现棒材穿水后温度控制在较合理的范围内,保证了棒材的高质量生产。通过与传统棒材穿水冷却控制系统以及分布式多处理器闭环控制系统的上位机监控显示轧后温度的稳定性分析,说明了基于分布式多处理器控制的棒材穿水冷却系统设计的可行性,不但使控制系统的稳定性和可靠性得到提高,同时还实现了控制系统的快速响应,减少了过多的人工干预,提升了棒材的冷却精度。

关键词： List-scheduling   Multiprocessor task-graph scheduling   Parallel and distributed systems   Task priority measurements  

摘要： Optimized task scheduling is one the most important factors to achieve high-performance in multiprocessor environments such as parallel and distributed systems. A large number of proposed approaches to solve this problem use list-scheduling technique in which a list of tasks is created based on the some priority measurements, and then in each step, the most priority task in the list is selected to schedule on the processor that allows the earliest start time. Therefore, the achieved schedule length is highly coherent with how order the tasks are selected to execute. Whereas selected task priority measurement determines which task order would be extracted, in this paper, we survey five traditional task priority measurements named height-level (HL), top-level (TL), bottom-level (BL), static-level (SL) and as-late-as-possible (ALAP) which have been extensively used in the different list-scheduling approaches. In addition, a new efficient task priority measurement based on the number-of-offspring (NOO) of tasks in task-graph is introduced and evaluated beside the others. The evaluation is made by doing various experiments on random task-graphs with different shape parameters and task-graphs of real-world programs using measures such as normalized schedule length (NSL), pair-wise and global comparison and best solution. Based on the results, it can generally be seen that the proposed NOO is the best, HL, BL, SL and ALAP are temperate (their ranks are slightly changed based on the selected comparison measure), and TL has the worst performance. © 2015 SERSC.

关键词： List-scheduling   Multiprocessor task-graph scheduling   Parallel and distributed systems   Task priority measurements  

摘要： One of the most important challenges to achieve high-performance in multiprocessor environments such as parallel and distributed systems is task scheduling. In such architectures, each program is decomposed into the smaller and dependent segments so-called tasks. To formulate the problem, execution times of the tasks, precedence constrains and communication costs among them are modeled using a directed acyclic graph (DAG) named task graph. The goal is to minimize the program completion-time (makespan) by means of mapping the tasks to a predefined number of identical processors in such a way that precedence constrains are preserved. This problem is shown to be NP-hard in general form, and hence, a number of heuristic approaches to solve it have been introduced. A large number of proposed approaches in the literature use list-scheduling technique in which a list of tasks is created based on some priority measurements, and then in each step, the most priority task in the list is selected to schedule on the processor that allows the earliest start time (EST) until all tasks are scheduled. Assigning tasks to the processors using the EST is based on the two different techniques: insertion-based assigning (IBA) and non-insertion based (non-IBA). In this paper, we survey these two different approaches in details, and analyze the strategies and philosophies behind them. In addition, a comprehensive set of experiments and evaluations from different points of view has been done, and various results and conclusions have been presented. © 2015 SERSC.

关键词： Mathematical modelling   Operating system   Processes planning   Queuing system   Service discipline   Stochastic network   Task manager  

摘要： The paper aims at carrying out the mathematical modeling and the performance analysis of a parallel computer system. Research methods are based on using the theory of analytical, numerical and simulation modeling, the theory of systems and queuing networks, the probability theory and the stochastic process theory. The authors deal with the analytical models of Task Managers for parallel processing systems based on the open-loop queuing systems. They investigate the methods of studying first-in-first-out Task Managers. The analytical models based on stochastic queuing networks for obtaining the managers' probabilistic and temporal characteristics are presented in the article. The results of the work done are equations for calculating the mean residence time of the problem in each system under study. The analytical calculations have been verified for their adequacy by simulation modeling. The experimental results have been displayed on the graph. During the problem investigation the appropriate conclusions have been made for each system type presented in the article. The considered models of Managers can be used in general purpose systems (for example in operating systems). © 2006-2015 Asian Research Publishing Network (ARPN).