Parallel implementation of the TRANSIMS micro-simulation

Abstract

This paper describes the parallel implementation of the TRANSIMS traffic micro-simulation. The parallelization method is domain decomposition, which means that each CPU of the parallel computer is responsible for a different geographical area of the simulated region. We describe how information between domains is exchanged, and how the transportation network graph is partitioned. An adaptive scheme is used to optimize load balancing.We then demonstrate how computing speeds of our parallel micro-simulations can be systematically predicted once the scenario and the computer architecture are known. This makes it possible, for example, to decide if a certain study is feasible with a certain computing budget, and how to invest that budget. The main ingredients of the prediction are knowledge about the parallel implementation of the micro-simulation, knowledge about the characteristics of the partitioning of the transportation network graph, and knowledge about the interaction of these quantities with the computer system. In particular, we investigate the differences between switched and non-switched topologies, and the effects of 10 Mbit, 100 Mbit, and Gbit Ethernet.As one example, we show that with a common technology -100 Mbit switched Ethernet -one can run the 20 000-link EMME/2-network for Portland (Oregon) more than 20 times faster than real time on 16 coupled Pentium CPUs.

Increased Convergence Rates in Multiagent Transport Simulations with Pseudosimulation

“…Although the remaining tasks of a single iteration are straightforwardly distributed across multiple computational nodes, this is not true for network loading. The reason is that the physical system is tightly integrated: if a vehicle reacts to another vehicle in a typical reaction time of 1 s, neighboring simulation items should not go out of synchronization by more than 1 s. In this situation, spatial decomposition minimizes interactions most (27) and may allow somewhat longer synchronization delays when network links are sufficiently long (28). However, parallel implementations of network loading are difficult to keep stable in terms of software engineering, and making them more stable affects performance (12).…”

Behavioral simulations in MapReduce

“…Behavioral simulations are often extremely complex, involving sophisticated numerical computations and elaborate decision procedures. Much existing work on behavioral simulations is from the high-performance computing community, and they frequently resort to hand-coding specific simulations in a low-level language to achieve acceptable performance [21,38]. A general purpose framework must be competitive with these handcoded applications in order to gain acceptance.…”

“…Given this situation, behavioral simulation developers have resorted to hand-coding parallel implementations of specific simulation models [21,38], or trading model accuracy for scalability and ease of implementation [8,50]. Our work, in contrast, aligns in spirit with recent efforts to bring dataflow programming techniques to complex domains, such as distributed systems [1] and networking [32], with huge benefits in performance and programming productivity.…”

Parallel implementation of a transportation network model

“…Nagel and Rickert [54] make use of the master-slave approach using decomposition for a parallel implementation of the TRANSIMS micro-simulation. Reference is also made to…”