Tobias Kiesling’s research while affiliated with Industrieanlagen-Betriebsgesellschaft and other places

Abstract

Queuing systems are an important building block for performance evaluation in various application areas, due to their powerful, yet simple nature. Although it is often possible to perform an analytical evaluation of a queuing model, simulation of queuing systems remains an important technique in the context of performance evaluation. In order to speed up queuing simulation executions, parallel and distributed simulation techniques have been devised. Unfortunately, existing methods are complex in nature, leading to increased development costs. Moreover, most of these approaches have been developed for tightly coupled parallel processing machines. Consequently, they are not suited for a distributed computing environment. This paper investigates an alternative approach based on the technique of time-parallel simulation with fix-up computations. The salient features of this novel approach are its simplicity and its suitability for execution in a distributed environment

Parallel simulation techniques are designed to increase simulation model performance by exploiting model concurrency. Unfortunately, designing efficient parallel simulations is not always an easy task. Most existing techniques guarantee results identical to a corresponding sequential simulation. Other methods try to increase efficiency by relaxing causal constraints, leading to the calculation of approximate results. This work proposes to combine both approaches using the novel technique of progressive time-parallel simulation, where imprecise results are calculated rapidly and improved progressively later on until precise results are known. The user is allowed to cancel this process at any time if the accuracy is satisfying. Progressive time-parallel simulation is a specialized parallelization approach that is not applicable to every kind of simulation model. Possible application areas include simulation-based decision support or simulation-based scheduling and control of manufacturing systems. A successful application of the technique is illustrated with progressive queuing system simulation.

Even after several decades of research, modeling is considered an art, with a high liability to produce incorrect abstractions of real world systems. Therefore, validation and verification of simulation models is considered an indispensable method to establish the credibility of developed models. In the process of parallelizing or distributing a given credible simulation model, a bias is introduced, possibly leading to serious errors in simulation results. Depending on the mechanisms used for parallelization or distribution, a separate validation of the parallel or distributed model is required. A necessary first step for such a validation is an understanding of the sources of bias that might occur through parallelization or distribution of a simulation model. The intention of this paper is to give an overview of the various types of bias and to give a formal definition of the bias and its quantification.

As an alternative to spatial parallelization of simulation models, time-parallel simulation offers the potential for massive parallelism with a high level of independence between the parallel processes. Unfortunately, due to inherent problems, the applicability of time-parallel simulation is restricted. Therefore, it has been proposed recently, to use approximation with time-parallel simulation in order to facilitate its application and to extend the class of models suitable for time-parallel simulation. As a proof-of-concept, this work shows how approximate temporal parallelization can be applied to the simulation of road traffic. Traffic simulation is used extensively in transportation research for various purposes, e.g. analysis of traffic phenomena, traffic forecast, and optimization of traffic flow. Depending on the level of fidelity, a traffic model exhibits a state space of moderate to high complexity. This paper is intended to discuss the basic properties of time-parallel traffic simulation and to examine its feasibility. Experiments with a sequential microscopic traffic simulator, that emulates important aspects of a corresponding time-parallel simulator, suggest this feasibility.

Temporal decomposition of simulation models, which is used in time-parallel simulation, is a promising alternative to classical spatial decomposition. It has been successfully applied to a small number of different models, most prominently for the simulation of computer caches and queuing systems. Unfortunately, widespread use is prevented by the state-match problem, which restricts the application of time-parallel simulation. Instead of a correct solution to the state-match problem, this work proposes the use of approximate solutions to facilitate the temporal decomposition of simulation models and to extend the class of models suitable for time-parallel simulation. However, this introduces an error in the simulation results, which might seriously distort or even invalidate results. Therefore, the error must be closely analyzed, and a method of error control must be provided. In addition to the basic properties of approximate techniques in time-parallel simulation, this work presents two use cases that illustrate the introduced concepts.

In time-parallel simulation, the simulation time axis is decomposed into a number of slices, which are assigned to parallel processes for concurrent simulation. Although a promising parallelization technique, it is difficult to be applied. Recently, using approximation with time-parallel simulation has been proposed to extend the class of suitable models and to improve the performance of existing models. In trace-driven cache simulation, sequences of memory requests are processed to determine the performance of variously sized caches. Time-parallel simulation has been applied to trace-driven cache simulation, but only with limited scalability of the parallel algorithm. In order to solve the scaling problem, this work uses approximation with time-parallel cache simulation. Although introducing an uncertainty in the results, the approximate algorithms work in a way that result accuracy increases monotonically with time, allowing direct control of the quality of results. Experiments with a prototypical implementation indicate the viability of this approach.

Time-parallel simulation offers the potential of massive parallelization of a simulation application, due to the amount of achievable parallelism not being restricted by the decomposability of the state space of a simulation model. Unfortunately, the potential speedup of a time-parallel simulation highly depends on the ability to match final and initial states of adjacent time intervals. Depending on the properties of the underlying simulation model, it might be feasible to accept a simulation iteration, even if the states of adjacent time intervals do not match exactly. This leads to the concept of approximative state matching in time-parallel simulation, which is introduced in this paper. Experiments with a prototypical implementation of a simple simulation model show encouraging results in terms of simulation speedup and introduced error.

Previous work on simulation of LRU caching led to the development of paral- lel algorithms that are efficient for small numbers of processors. However, these algorithms exhibit a sub-linear speedup, where the efficiency seriously decreases with a higher number of processors. In order to achieve linear speedup, this work proposes the use of approximation techniques with the existing parallelization ap- proaches. The nature of the approximate algorithm allows direct control of the introduced error, which can be used to achieve reasonable speedup with a minimal error.

Queuing systems are an important building block for per-formance evaluation in various application areas, due to their powerful, yet simple nature. Although it is often possible to perform an analytical evaluation of a queuing model, simulation of queuing systems remains an important technique in the context of performance evaluation. In order to speed up queuing simulation executions, parallel and distributed sim-ulation techniques have been devised. Unfortunately, existing methods are complex in nature, leading to increased development costs. More-over, most of these approaches have been developed for tightly coupled parallel processing machines. Consequently, they are not suited for dis-tributed execution. This paper investigates an alternative approach based on the technique of time-parallel simulation with fix-up computations. The salient features of this novel approach are its simplicity and its suit-ability for efficient distributed simulation execution.