Spatiotemporal Domain Decomposition for Massive Parallel Computation of Space-Time Kernel Density

Abstract

ABSTRACT:Accelerated processing capabilities are deemed critical when conducting analysis on spatiotemporal datasets of increasing size, diversity and availability. High-performance parallel computing offers the capacity to solve computationally demanding problems in a limited timeframe, but likewise poses the challenge of preventing processing inefficiency due to workload imbalance between computing resources. Therefore, when designing new algorithms capable of implementing parallel strategies, careful spatiotemporal domain decomposition is necessary to account for heterogeneity in the data. In this study, we perform octtree-based adaptive decomposition of the spatiotemporal domain for parallel computation of space-time kernel density. In order to avoid edge effects near subdomain boundaries, we establish spatiotemporal buffers to include adjacent data-points that are within the spatial and temporal kernel bandwidths. Then, we quantify computational intensity of each subdomain to balance workloads among processors. We illustrate the benefits of our methodology using a space-time epidemiological dataset of Dengue fever, an infectious vector-borne disease that poses a severe threat to communities in tropical climates. Our parallel implementation of kernel density reaches substantial speedup compared to sequential processing, and achieves high levels of workload balance among processors due to great accuracy in quantifying computational intensity. Our approach is portable of other space-time analytical tests.

Space-Time Kernel Density Estimation for Real-Time Interactive Visual Analytics

“…Dengue Fever Data. As part of our results we compared our algorithm to that of prior work on Dengue fever outbreaks in Cali, Columbia [14]. The dengue fever dataset is important because it is based on real world cases of dengue fever.…”

“…As a result of these efforts, we were able to achieve great performance. We used a Dengue Fever dataset [14]. We initially tested two different configurations: one using just the CPU and one using just the GPU.…”

“…We acquired a copy of the Dengue Fever data from prior work to evaluate our algorithm [14]. For our comparisons we used the same data, the same temporal, spatial and voxel size in order to maintain control on all these metrics.…”

“…Running the STKDE sequentially takes approximately 23 minutes. Hohl et al [14] proposed a decomposition step that must be performed before running the STKDE. For this setup the decomposition step took 62.96 seconds.…”

“…Recently, there have been several prior works that focus on speeding up the KDE. Hohl et al [14] have used a parallel CPU approach to compute the STKDE. In their method they use a spatial decomposition initially to subdivide the work load among a defined set of CPUs.…”