一种面向移动Agent的多任务并行计算模型及算法

介绍了一种面向移动Agent的并行计算模型,给出了采用十标度策略解决任务排序,采用满射策略解决任务映射的算法。该模型允许多个计算任务在异构主机构成的分布式环境下同时进行计算,并且通过算法优化,降低移动Agent之间的通信成本,减少网络流量。     

分布式流体系结构及其编程模型与资源管理

利用互联网资源提供大数据计算服务面临着资源异构性、动态性与通信长延迟等方面的挑战,现有分布式计算模型仍存在一些不足。运用流计算模型提出分布式流体系结构,包括分布式流编程模型与资源管理等,能够高效支持多种并行执行模式。在10个CPU-GPU异构结点上实现了原型系统,仿真实验验证了7个不同的测试用例。实验结果表明,与本地串行计算相比,分布式流体系结构可以平均提高39倍计算性能,具有较大的应用潜力。     

Optimal joint preventive maintenance and production policies

We study joint preventive maintenance (PM) and production policies for an unreliable production‐inventory system in which maintenance/repair times are non‐negligible and stochastic. A joint policy decides (a) whether or not to perform PM and (b) if PM is not performed, then how much to produce. We consider a discrete‐time system, formulating the problem as a Markov decision process (MDP) model. The focus of the work is on the structural properties of optimal joint policies, given the system state comprised of the system's age and the inventory level. Although our analysis indicates that the structure of optimal joint policies is very complex in general, we are able to characterize several properties regarding PM and production, including optimal production/maintenance actions under backlogging and high inventory levels, and conditions under which the PM portion of the joint policy has a control‐limit structure. In further special cases, such as when PM set‐up costs are negligible compared to PM times, we are able to establish some additional structural properties. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005.     

The continuous‐time single‐sourcing problem with capacity expansion opportunities

This paper develops a new model for allocating demand from retailers (or customers) to a set of production/storage facilities. A producer manufactures a product in multiple production facilities, and faces demand from a set of retailers. The objective is to decide which of the production facilities should satisfy each retailer's demand, in order minimize total production, inventory holding, and assignment costs (where the latter may include, for instance, variable production costs and transportation costs). Demand occurs continuously in time at a deterministic rate at each retailer, while each production facility faces fixed‐charge production costs and linear holding costs. We first consider an uncapacitated model, which we generalize to allow for production or storage capacities. We then explore situations with capacity expansion opportunities. Our solution approach employs a column generation procedure, as well as greedy and local improvement heuristic approaches. A broad class of randomly generated test problems demonstrates that these heuristics find high quality solutions for this large‐scale cross‐facility planning problem using a modest amount of computation time. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005.     

Performance evaluation of a production/inventory system with periodic review and endogenous lead times

This paper considers a discrete time, single item production/inventory system with random period demands. Inventory levels are reviewed periodically and managed using a base‐stock policy. Replenishment orders are placed with the production system which is capacitated in the sense that there is a single server that sequentially processes the items one at a time with stochastic unit processing times. In this setting the variability in demand determines the arrival pattern of production orders at the queue, influencing supply lead times. In addition, the inventory behavior is impacted by the correlation between demand and lead times: a large demand size corresponds to a long lead time, depleting the inventory longer. The contribution of this paper is threefold. First, we present an exact procedure based on matrix‐analytic techniques for computing the replenishment lead time distribution given an arbitrary discrete demand distribution. Second, we numerically characterize the distribution of inventory levels, and various other performance measures such as fill rate, base‐stock levels and optimal safety stocks, taking the correlation between demand and lead times into account. Third, we develop an algorithm to fit the first two moments of the demand and service time distribution to a discrete phase‐type distribution with a minimal number of phases. This provides a practical tool to analyze the effect of demand variability, as measured by its coefficient of variation, on system performance. We also show that our model is more appropriate than some existing models of capacitated systems in discrete time. © 2007 Wiley Periodicals, Inc. Naval Research Logistics, 2007     

Non‐greedy heuristics and augmented neural networks for the open‐shop scheduling problem

In this paper we propose some non‐greedy heuristics and develop an Augmented‐Neural‐Network (AugNN) formulation for solving the classical open‐shop scheduling problem (OSSP). AugNN is a neural network based meta‐heuristic approach that allows integration of domain‐specific knowledge. The OSSP is framed as a neural network with multiple layers of jobs and machines. Input, output and activation functions are designed to enforce the problem constraints and embed known heuristics to generate a good feasible solution fast. Suitable learning strategies are applied to obtain better neighborhood solutions iteratively. The new heuristics and the AugNN formulation are tested on several benchmark problem instances in the literature and on some new problem instances generated in this study. The results are very competitive with other meta‐heuristic approaches, both in terms of solution quality and computational times. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005.     

适用于云计算的多接收方代理重加密方案

在云计算环境下,绝大多数信息处理和存储都被转移到云中心,用户终端的计算和存储能力将逐渐弱化。本文针对这一趋势,提出一种适用于云计算的多接收方代理重加密方案,充分利用云中心的分布式计算能力,使用随机数重用和密文聚合技术,大大降低了计算和传输负担,实验表明本方案具有较高的效率。     

Outsourcing warranty repairs: Dynamic allocation

In this paper we consider the problem of minimizing the costs of outsourcing warranty repairs when failed items are dynamically routed to one of several service vendors. In our model, the manufacturer incurs a repair cost each time an item needs repair and also incurs a goodwill cost while an item is awaiting and undergoing repair. For a large manufacturer with annual warranty costs in the tens of millions of dollars, even a small relative cost reduction from the use of dynamic (rather than static) allocation may be practically significant. However, due to the size of the state space, the resulting dynamic programming problem is not exactly solvable in practice. Furthermore, standard routing heuristics, such as join‐the‐shortest‐queue, are simply not good enough to identify potential cost savings of any significance. We use two different approaches to develop effective, simply structured index policies for the dynamic allocation problem. The first uses dynamic programming policy improvement while the second deploys Whittle's proposal for restless bandits. The closed form indices concerned are new and the policies sufficiently close to optimal to provide cost savings over static allocation. All results of this paper are demonstrated using a simulation study. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005     

Disruption management in production planning

We study the problem of recovering a production plan after a disruption, where the disruption may be caused by incidents such as power failure, market change, machine breakdown, supply shortage, worker no‐show, and others. The new recovery plan we seek after has to not only suit the changed environment brought about by the disruption, but also be close to the initial plan so as not to cause too much customer unsatisfaction or inconvenience for current‐stage and downstream operations. For the general‐cost case, we propose a dynamic programming method for the problem. For the convex‐cost case, a general problem which involves both cost and demand disruptions can be solved by considering the cost disruption first and then the demand disruption. We find that a pure demand disruption is easy to handle; and for a pure cost disruption, we propose a greedy method which is provably efficient. Our computational studies also reveal insights that will be helpful to managing disruptions in production planning. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005.     

适应低时延业务需求的分布式可穿戴单兵作战信息系统

针对云计算应用在单兵作战系统场景下业务处理时延高、服务质量无法保障的问题,提出一种基于可穿戴计算的分布式单兵作战信息系统。利用士兵身上的可穿戴智能设备构建本地计算层,在作战地点就近处理计算数据,提供给士兵本地的信息处理与融合能力,并采用广义扩散负载均衡算法平衡各设备负载,降低业务处理时延;同时利用分布式计算的容错能力增强系统的可靠性。仿真结果表明,基于可穿戴设备的分布式本地网络架构能有效地降低作战任务的处理时延,同时增强系统的可靠性。     

移动智能体的形式化研究

移动智能体是当前分布计算领域的研究热点之一 ,但是其理论研究还处于一个很不成熟的阶段。文中阐述了对移动智能体作形式化研究的必要性 ,介绍了当前的研究现状 ,分析了其中的不足 ,并进一步给出了对后续工作的展望     

Analysis of a production system in a general configuration

We consider a production system comprising multiple stations (or workshops) such as an entry station, a set of work stations, a central station, and an exit station, which are arranged in a general configuration. A worker (or a vehicle tool) is assigned to each station, who sends a part from the station to the destination station according to the required process path of the part. Any part is allowed to visit a work station more than once if its process path requires. We propose a new control strategy with the push policy for instructing each worker to send a part and the kanban mechanism for controlling the work‐in‐process (WIP) in each work station. As all work stations have limited local buffers, the central station is used for storing blocked parts temporarily. Such a production system is modeled as an open queueing network in a general configuration with a Markovian part sending policy and a machine no blocking mechanism. The queueing network is analytically characterized. Some important performance measures are compared with other control strategies. A semi‐open decomposition approach is applied to the queueing network for computing the blocking probabilities when parts arrive at the work stations. An algorithm is developed based on the semi‐open decomposition approach. Numerical experiments show the quality of the solutions obtained by the algorithm as well as a property of a performance measure. © 2001 John Wiley & Sons, Inc. Naval Research Logistics 48: 128–143, 2001     

差分离散方法分布式并行计算的重叠边界优化

针对分布式存储环境下显式差分方法的并行计算问题 ,依据分布式存储多处理机的体系结构特点 ,提出了一个重叠边界优化模型 ;该模型目前已成功应用于中国科学院第三代海洋环流模式的分布式并行计算优化。     

高性能异构加速器MiniGo算子优化方法

根据高性能异构加速器的特性和MiniGo的训练模式提出了一种高效的并行计算方法。对片上计算资源进行合理规划,实现异构设备之间的流水并行优化;根据异构设备间存在共享存储段设计了共享内存编码模式,减少数据传输开销;根据数字信号处理簇内具有多计算资源的特点结合算子计算-访存特性设计了不同的算子并行计算优化策略。同时,面向TensorFlow实现了一个易于使用的高性能计算库。实验结果显示,该方法实现了典型算子的多核并行计算。相对于单核,卷积算子加速比为24.69。相较于裁剪版8核FT2000+CPU,该方法训练和自博弈执行速度加速比分别为3.83和1.5。     

Capacity allocation with machine duplication in semiconductor manufacturing

The manufacturing process for a computer chip is complex in that it involves a large number of distinct operations requiring a substantial lead‐time for completion. Our observations of such a manufacturing process at a large plant in the United States led us to identify several tactical and operational problems that were being addressed by the production planners on a recurring basis. This paper focuses on one such problem. At a tactical level, given a demand forecast of wafers to be manufactured, one specific problem deals with specifying which machine or machine groups will process different batches of wafers. We address this problem by recognizing the capacity limitations of the individual machines as well as the requirement for reducing operating and investment costs related to the machines. A mathematical model, which is a variation of the well‐known capacitated facility location problem, is proposed to solve this problem. Given the intractability of the model, we first develop problem specific lower bounding procedures based on Lagrangean relaxation. We also propose a heuristic method to obtain “good” solutions with reasonable computational effort. Computational tests, using hypothetical and industry‐based data, indicate that our heuristic approach provides optimal/near optimal solutions fairly quickly. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005     

Dynamic policy for idling time preservation

This study aims to determine and evaluate dynamic idling policies where an agent can idle while some customers remain waiting. This type of policies can be employed in situations where the flow of urgent customers does not allow the agent to spend sufficient time on back-office tasks. We model the system as a single-agent exponential queue with abandonment. The objective is to minimize the system's congestion while ensuring a certain proportion of idling time for the agent. Using a Markov decision process approach, we prove that the optimal policy is a threshold policy according to which the agent should idle above (below) a certain threshold on the queue length if the congestion-related performance measure is concave (convex) with respect to the number of customers present. We subsequently obtain the stationary probabilities, performance measures, and idling time duration, expressed using complex integrals. We show how these integrals can be numerically computed and provide simpler expressions for fast-agent and heavy-traffic asymptotic cases. In practice, the most common way to regulate congestion is to control access to the service by rejecting some customers upon arrival. Our analysis reveals that idling policies allow high levels of idling probability that such rejection policies cannot reach. Furthermore, the greatest benefit of implementing an optimal idling policy occurs when the objective occupation rate is close to 50% in highly congested situations.     

A generalization of the weighted set covering problem

We study a generalization of the weighted set covering problem where every element needs to be covered multiple times. When no set contains more than two elements, we can solve the problem in polynomial time by solving a corresponding weighted perfect b‐matching problem. In general, we may use a polynomial‐time greedy heuristic similar to the one for the classical weighted set covering problem studied by D.S. Johnson [Approximation algorithms for combinatorial problems, J Comput Syst Sci 9 (1974), 256–278], L. Lovasz [On the ratio of optimal integral and fractional covers, Discrete Math 13 (1975), 383–390], and V. Chvatal [A greedy heuristic for the set‐covering problem, Math Oper Res 4(3) (1979), 233–235] to get an approximate solution for the problem. We find a worst‐case bound for the heuristic similar to that for the classical problem. In addition, we introduce a general type of probability distribution for the population of the problem instances and prove that the greedy heuristic is asymptotically optimal for instances drawn from such a distribution. We also conduct computational studies to compare solutions resulting from running the heuristic and from running the commercial integer programming solver CPLEX on problem instances drawn from a more specific type of distribution. The results clearly exemplify benefits of using the greedy heuristic when problem instances are large. © 2003 Wiley Periodicals, Inc. Naval Research Logistics, 2005     

Stochastic analysis of a single server retrial queue with general retrial times

Retrial queueing systems are widely used in teletraffic theory and computer and communication networks. Although there has been a rapid growth in the literature on retrial queueing systems, the research on retrial queues with nonexponential retrial times is very limited. This paper is concerned with the analytical treatment of an M/G/1 retrial queue with general retrial times. Our queueing model is different from most single server retrial queueing models in several respectives. First, customers who find the server busy are queued in the orbit in accordance with an FCFS (first‐come‐first‐served) discipline and only the customer at the head of the queue is allowed for access to the server. Besides, a retrial time begins (if applicable) only when the server completes a service rather upon a service attempt failure. We carry out an extensive analysis of the queue, including a necessary and sufficient condition for the system to be stable, the steady state distribution of the server state and the orbit length, the waiting time distribution, the busy period, and other related quantities. Finally, we study the joint distribution of the server state and the orbit length in non‐stationary regime. © 1999 John Wiley & Sons, Inc. Naval Research Logistics 46: 561–581, 1999     

基于R-树索引的Map-Reduce空间连接聚集操作

空间连接聚集是一种常用并且非常耗时的空间数据库操作,特别是在面对大规模空间数据集时,单机运行环境难以满足其对时空开销的需求,如何设计高效的面向云计算环境中的分布式空间连接聚集算法越来越受到人们关注。Map-Reduce作为云计算的核心模式受限于其扁平化的串行扫描操作模型,常被用来加速非索引的空间连接操作,现有工作尚无将Map-Reduce和R-树索引结合来处理空间连接聚集。因此,提出了基于R-树索引的Map-Reduce空间连接聚集算法(RSJA-MR)来更高效地返回连接聚集结果。提出一种分布式R-树索引结构以支持大规模空间数据的索引,RSJA-MR算法利用分布式R-树生成任务集,任务集的执行满足无依赖并行计算模式,很容易在Map-Reduce框架中进行表达。文中提出一种实时缓存策略以支持索引并发访问。实验结果表明:相比非索引的Map-Reduce连接聚集算法,在空间交叠连接聚集查询上,时间性能最少提升8%,在空间包含连接聚集查询上,时间性能最少提升近35%。     

量子算法模拟实现技术研究

量子计算并行性展示了强大的运算能力 ,但现实条件决定了目前量子算法研究主要依靠模拟进行。本文介绍了量子计算的基本原理和计算模型 ,对量子算法的模拟实现技术进行了分析。通过对量子算法模拟的时空分析 ,指出目前量子算法模拟中存在的一些问题 ,讨论了今后的一些研究方向。