Frank Thilo

A comparative study of online scheduling algorithms for networks of workstations

Networks of workstations offer large amounts of unused processing time. Resource management systems are able to exploit this computing capacity by assigning compute-intensive tasks to idle workstations. To avoid interferences between multiple, concurrently running applications, such resource management systems have to schedule application jobs carefully. Continuously arriving jobs and dynamically changing amounts of available CPU capacity make traditional scheduling algorithms difficult to apply in workstation networks. Online scheduling algorithms promise better results by adapting schedules to changing situations. This paper compares six online scheduling algorithms by simulating several workload scenarios. Based on the insights gained by simulation, the three online scheduling algorithms performing best were implemented in the Winner resource management system. Experiments conducted with Winner in a real workstation network confirm the simulation results obtained.

Load distribution in a CORBA environment

The design and implementation of a CORBA load distribution service for distributed scientific computing applications running in a network of workstations is described. The proposed approach is based on integrating load distribution into the CORBA naming service which in turn relies on information provided by the underlying WINNER resource management system developed for typical networked Unix workstation environments. The necessary extensions to the naming service, the WINNER features for collecting load information and the placement decisions are described. A prototypical implementation of the complete system is presented, and performance results obtained for the parallel optimization of a mathematical test function are discussed.

Distributed solution of optimal hybrid control problems on networks of workstations

The design of an optimal control strategy for a hybrid system is a matter of growing interest in computational engineering. The solution of optimization problems in most engineering disciplines often requires efficient parallel optimization algorithms to solve these kinds of problems in reasonable time. Instead of introducing parallelism to selected components of an existing sequential algorithm, the algorithm proposed in this paper is aimed at utilizing the available computational resources efficiently throughout the course of the optimization. To assure a certain level of efficiency the algorithm can be adapted to the available resources and the dimension of the problem to be solved. The features of this inherently parallel algorithm are described, and the parallel performance is analyzed by means of a scalability analysis. To demonstrate the use of the algorithm for the solution of optimization problems in computational engineering, two problems from groundwater engineering are solved.

Grid -based Computing for Multidisciplinary Analysis and Optimization

Multidisciplinary design demands the simulation of complex systems in order to evaluate the behaviour of these systems under different objectives and control strategies. Finding a “good” – or even maybe an optimal – design for a system implies the solution of an optimization problem based on the results of typically hundreds to thousands simulations. The application of computationally expensive simulations in the course of such multidisciplinary optimizations results in long -running solution processes even when using state -of -the -art parallel/distributed algorithms and hardware. The grid -based solution of this kind of optimization problems demands certain features of parallel/distributed systems: efficient utilization of resources (i.e. processors), adequate optimization algorithms (i.e. inherently parallel optimization algorithms), and software integration (i.e. integration of optimization algorithms and simulation code). In this paper, multidisciplinary optimization tasks are characterized and a grid -based problem solving environment with a corresponding scalable algorithm is presented. The usability of t he approach is demonstrated by applying it to different problems from groundwater engineering, automotive industry and airplane design.

On Scalability of Distributed Simulation and Optimization of Forming Processes

Advances in the processing power of modern computer hardware and the availability of sophisticated simulation software have fostered the use of virtual prototyping techniques in the automotive industry. However, many engineering problems require simulation and optimization computations so complex that it is imperative to utilize parallel computing to solve these problems in reasonable time. In this paper, FETI‐INDEED — a parallelized simulation code for sheet metal forming — as well as algorithms for distributed optimizations are examined. Because it is anticipated that increasingly large problems will need to be solved in the future, particular focus is put on scalability aspects of these components. Experimental results are provided for optimizations which were performed on a 300 CPU Opteron cluster.

Verteilte Lösung simulationsbasierter Optimierungsprobleme auf vernetzten Workstations

In diesem Beitrag wird eine Architektur fur eine Softwareumgebung vorgestellt, die die verteilte Losung typischer Klassen simulationsbasierter ingenieurwissenschaftlicher Optimierungsprobleme auf vernetzten Workstations unterstutzt. Die Anforderungen der softwaretechnischen Kopplung von Simulations- und Optimierungssoftware und deren Verteilung werden diskutiert, und anhand dieser Anforderungen wird eine adaquate Architektur entwickelt. Entwurf und Implementierung der Hauptkomponenten — die Simulations- und Optimierungskomponente — sowie deren Integration in diese Architektur werden naher vorgestellt. Auf der Simulationsseite wird dabei ein kommerzielles Finite-Elemente Simulationssystem fur Stromungs- und Transportprozesse im Grundwasser eingesetzt. Der Algorithmus auf der Optimierungsseite ist insbesondere fur die Losung simulations-basierter nichtlinearer Optimierungsprobleme unter Nebenbedingungen entworfen und implementiert worden. Mit dem vorgestellten Algorithmus werden Problemstellungen aus dem Grundwassermanagement gelost.

Grid Computing for Optimization and Sensitivity Analysis in Groundwater Engineering

In the field of groundwater engineering complex models are used to simulate environmental processes and to enable measurement planning. To improve model quality and reliability and to enable more efficient environmental planning often large numbers of model runs are required during different model lifecycles. The paper describes an approach how to use grid technology to speedup model design as well as operational tasks like modelbased optimization. Beside the background of groundwater engineering problems, a Globus Toolkit based architecture to solve such tasks is introduced and initial results are presented. I. Introduction EGARDING the field of numerical groundwater modeling (typically based on Finite-Element methods), increasingly large areas and / or complicated hydrogeological structures have to be simulated and are often represented by complex 3-dimensional FEM meshes. Concurrently the performance increase of traditional single CPU systems is beginning to decelerate. Therefore, new techniques have to be used to solve complex groundwater modeling tasks on parallel computing resources. Two principal ways to speed up groundwater simulation can be used: applying fine-grained parallelism in the core of the simulation software itself or to execute several simulations simultaneously, thus utilizing parallelism on a more coarse-grained level. The former can be used to potentially speed up any simulation, but requires changes to the solver code. The latter can be applied, whenever several simulations would have to be run consecutively in a traditional, sequential setup and the input of one simulation does not depend on the result of another. This paper will concentrate on the second parallelization approach. During the groundwater modeling workflow, series of model applications are common at different lifecycles of the model. These include the calibration of model parameters, sensitivity analysis of parameters which have to be estimated, and technical model based optimization. A typical example for the last application field is to minimize operational costs of pumping stations while still satisfying constraints like a minimum allowed depth to water table or to meet certain groundwater quality thresholds. Each of the above tasks can lead to a large number of model simulation runs, which require no interconnects apart from modifying the parameters and returning the simulation results. Hence, these tasks are perfectly suited to be executed in parallel. In the past, mainly parallel computing in groundwater modeling was rare. If it took place, networks of workstations or compute clusters were the architecture of choice. The main disadvantage of these cluster-based

Parallel Direct Search Methods for Optimization of Forming Processes in the Automotive Industry

In many engineering disciplines, the use of realistic computing models has become an invaluable tool in the design process. Complex simulation codes are able to approximate the behaviour of intricate systems or the properties of components without the need for costly physical experimentation. Optimization algorithms can be used to automatically nd the set of parameters within the design space for which the simulation promises the most desirable characteristics. Parallel computing can help to get results within a feasible time in spite of the high computational demand of the simulations. A class of algorithms which can be successfully applied to this kind of problems are so-called parallel direct search methods. In this paper, several such search methods are described and computational results for optimizing a hydroforming process are presented.

Scalability Aspects of Distributed Simulation -based Optimization in Manufacturing

Advances in processing power of modern computer hardware allow the analysis of increasingly c omplex systems by means of simulation. However, many multidisciplinary engineering problems such as design optimization problems in the aircraft or automotive industry have extremely high computational demands. To solve these problems in reasonable time, t he computation has to be distributed over many CPUs, e.g. in form of a compute cluster or a computational grid. To exploit the total processing power of hundreds of CPUs, intelligent, scalable algorithms are required. In this paper, the concept of scalabil ity is examined in the context of parallel simulation software systems and the distributed solution of simulation -based optimization problems in manufacturing. As case studies, the optimization of alloy casting processes amd the simulation of metal -sheet f orming are presented. I. Introduction PTIMIZATION tasks in multidisciplinary optimization include design problems in manufacturing in the aircraft or automotive industry, alloy casting processes and metal -sheet forming. Typically, solving these problems inv olves running many complex simulations which are implemented in commercial software packages. In general, the optimization algorithm has to treat the simulation system as a black box. To solve such a simulation -based optimization problem, many hundreds or thousands of simulations are necessary, each of which is computationally expensive. This results in extremely high computational demands which can only be met by distributing the computation over many CPUs. Basically, there are two different approaches ho w parallel or distributed computing can be used in this scenario: First, the time needed for a single simulation run can be reduced by parallelizing the simulation software itself, e.g. by partitioning a FEM mesh and assigning each partition to a different CPU. Second, the optimization algorithm can request the evaluation of multiple scenarios at the same time, so that many (sequential) simulations are executed simultaneously. It is also possible to combine the two approaches. The simulation and optimizatio n of complex products in the area of virtual prototyping aims at reducing the time to market for new innovative products and creates an ever -increasing demand for computational power which can only be met by utilizing larger numbers of CPUs. This requires scalable hardware architectures and algorithms. This paper focuses on algorithmic scalability; in particular, three scalable optimization algorithms are presented. To allow a meaningful scalability analysis, a suitable performance and efficiency metric for heuristic, parallel optimization is developed. This is then used to analyze the characteristics of the algorithms for solving a benchmark problem. Further results are shown for computations of industrial pr oblems from casting processes

CORBA Based Runtime Support for Load Distribution and Fault Tolerance

Parallel scientific computing in a distributed computing environment based on CORBA requires additional services not (yet) included in the CORBA specification: load distribution and fault tolerance. Both of them are essential for long running applications with high computational demands as in the case of computational engineering applications. The proposed approach for providing these services is based on integrating load distribution into the CORBA naming service which in turn relies on information provided by the underlying WINNER resource management system developed for typical networked Unix workstation environments. The support of fault tolerance is based on error detection and backward reco very by introducing proxy objects which manage checkpointing and restart of services in case of failures. A protoytpical implementation of the complete system is presented, and performance results obtained for the parallel optimization of a mathematical benchmark function are discussed.