Profile-based spatial partitioning for parallel simulation of large-scale wildfires

Abstract

Spatial partitioning is commonly used for parallel simulation of spatial–temporal systems, such as simulations of wildfires. Achieving effective spatial partitioning is a challenging task due to the dynamic behavior of the simulation models. This paper presents a partitioning method named profile-based spatial partitioning for parallel simulation of large scale wildfires. The profile-based partitioning exploits the dynamic behavior of a wildfire spread simulation model and uses it as a profile to guide the spatial partitioning for parallel simulations. Experimental results show that the profile-based partitioning can increase the degree of parallelism and improve the scalability of parallel simulations of large-scale wildfires.

Introduction

Recent years have witnessed an increasing number of wildfires with increased burned areas as well as damages to the environment. Understanding the behavior of wildfires and predicting the fire intensity, rate of spread, and spread direction can support better management and control of wildfires. Towards this goal, several wildfire spread simulation models have been developed, including FARSITE [15], BehavePlus [14], Hfire [34], and DEVS-FIRE [9], [12]. Simulating large-scale wildfires is a computation-challenging task because of the complexity of the fire spread model and the large size and long duration of the simulations. To improve the performance of wildfire simulations, parallel simulation techniques are needed.

To support parallel simulations we need to divide the simulation tasks and assign them to multiple processing units (PUs). This step of task partitioning is critical and can have significant impact on the performance of a parallel simulation. Wildfire simulation models are characterized by their dynamic behavior in both space and time. Due to the spatial–temporal behavior, spatial partitioning is commonly employed for parallel simulations of wildfires. In spatial partitioning, the space is divided into multiple regions, which are distributed to multiple PUs. When a fire spreads to a new region, it triggers the computation of the PU in charge of that region. This paper considers spatial partitioning for parallel simulation of wildfire spread using the DEVS-FIRE model. DEVS-FIRE is a discrete event wildfire spread and suppression model based on the DEVS formalism [2]. It employs a two dimensional cellular space model composed from multiple individual cells to represent the wildfire area. When simulating large-scale wildfires modeled by a large number of cells, there is a need to improve the simulation performance using parallel simulation techniques.

An effective spatial partitioning for parallel simulation of wildfires should increase the degree of parallelism and improve the scalability as well. For the DEVS-FIRE model, one approach of spatial partitioning is to divide the cell space into several size-equivalent sub cell spaces corresponding to the number of PUs, and then allocate each sub cell space (also called a partition) to a PU for parallel simulation [22]. However, this approach is ineffective because the execution sequence of the simulation tasks heavily depends on the fire spreading behavior, which is influenced by many factors such as ignition point, vegetation, terrain, and weather [23]. The dynamic fire spreading behavior can cause different PUs to have significantly different computation load at a given time. To provide a simple example, let us consider a simulation where a wildfire is ignited at a location far from the partitioning boundary. In this case, it will take a long time for the fire to spread to other partitions to trigger them to start the simulation. As a result, in the beginning of the simulation only one PU containing the ignition point engages in computation; while others are idle waiting for the fire to spread to their partitions. This type of temporally unbalanced load of computation degrades the simulation performance.

In order to overcome the problem mentioned above, this paper presents a profile-based spatial partitioning method to improve the performance of large scale simulation of wildfires. The profile-based spatial partitioning is motivated from the observation of fire spreading behavior in DEVS-FIRE: the simulation focuses on the burning frontline of the fire that is the active area; no computation is needed for the area that is not ignited or already burned out. Thus if all PUs can concentrate on the active area of the fire, the time of waiting will be greatly reduced so that better parallelism can be achieved. Motivated by this, the profile-based partitioning method exploits the model behavior, i.e., the spatial–temporal wildfire spreading behavior, and uses it as a profile for guiding the task partitioning for parallel simulation. In our work, the behavior profile is obtained from a simulation using low resolution GIS data. The ignition sequence of the cells from the low resolution simulation represents a behavior profile of the high resolution simulation. This profile is used to partition the cell space for parallel simulation of DEVS-FIRE. A low resolution simulation runs much faster than the high resolution simulation. As will be shown later, the time spent in generating the profile (using the low resolution simulation) can be easily compensated by the performance gains from the profile-based partitioning. In this paper, the parallel simulation algorithm that we employed is based on the time warp algorithm [10], which is an optimistic approach of time management [26] that uses a rollback mechanism to recover violations of local causality in order to maintain the local causality constraint. We note that an alternative approach of parallel simulations is based on conservative algorithms (see e.g., [6]). Conservative algorithms rely on the concept of lookahead in order to achieve concurrent processing of events. The wildfire simulation model considered in this paper does not have good lookahead properties because its fire spreading behavior depends on the dynamic changing wind speed/direction as well as the non-uniform slope, aspect, and fuel model of the forest cells. Also note that in this paper, we only consider static partitioning where the partitions are divided and assigned to PUs before a simulation starts.

The remainder of this paper is organized as follows. Section 2 reviews the related work in task partitioning. Section 3 gives an overview of the DEVS-FIRE model. Section 4 presents the profile-based spatial partitioning method for parallel simulation of large scale wildfire. Section 5 gives a theoretical performance analysis of the profile-based method. Section 6 provides experiment results and Section 7 concludes this work.