Resumen de Novel parallel approaches to efficiently solve spatial problems on heterogeneous CPU-GPU systems
Andrés Jesús Sánchez Fernández
In recent years, approaches that seek to extract valuable information from large datasets have become particularly relevant in today's society. In this category, we can highlight those problems that comprise data analysis distributed across two-dimensional scenarios called spatial problems. These usually involve processing (i) a series of features distributed across a given plane or (ii) a matrix of values where each cell corresponds to a point on the plane. Therefore, we can see the open-ended and complex nature of spatial problems, but it also leaves room for imagination to be applied in the search for new solutions.
One of the main complications we encounter when dealing with spatial problems is that they are very computationally intensive, typically taking a long time to produce the desired result. This drawback is also an opportunity to use heterogeneous systems to address spatial problems more efficiently. Heterogeneous systems give the developer greater freedom to speed up suitable algorithms by increasing the parallel programming options available, making it possible for different parts of a program to run on the dedicated hardware that suits them best.
Several of the spatial problems that have not been optimised for heterogeneous systems cover very diverse areas that seem vastly different at first sight. However, they are closely related due to common data processing requirements, making them suitable for using dedicated hardware. In particular, this thesis provides new parallel approaches to tackle the following three crucial spatial problems: latent fingerprint identification, total viewshed computation, and path planning based on maximising visibility in large regions.
Latent fingerprint identification is one of the essential identification procedures in criminal investigations. Addressing this task is difficult as (i) it requires analysing large databases in a short time, and (ii) it is commonly addressed by combining different methods with complex data dependencies, making it challenging to exploit parallelism on heterogeneous CPU-GPU systems. Moreover, most efforts in this context focus on improving the accuracy of the approaches and neglect reducing the processing time—the most accurate algorithm was designed to process the fingerprints using a single thread. We developed a new methodology to address the latent fingerprint identification problem called Asynchronous processing for Latent Fingerprint Identification (ALFI) that speeds up processing while maintaining high accuracy. ALFI exploits all the resources of CPU-GPU systems using asynchronous processing and fine-coarse parallelism to analyse massive fingerprint databases. We assessed the performance of ALFI on Linux and Windows operating systems using the well-known NIST/FVC databases. Experimental results revealed that ALFI is on average 22x faster than the state-of-the-art identification algorithm, reaching a speed-up of 44.7x for the best-studied case.
In terrain analysis, Digital Elevation Models (DEMs) are relevant datasets used as input to those algorithms that typically sweep the terrain to analyse its main topological features such as visibility, elevation, and slope. The most challenging computation related to this topic is the total viewshed problem. It involves computing the viewshed—the visible area of the terrain—for each of the points in the DEM. The algorithms intended to solve this problem require many memory accesses to 2D arrays, which, despite being regular, lead to poor data locality in memory. We proposed a methodology called skewed Digital Elevation Model (sDEM) that substantially improves the locality of memory accesses and exploits the inherent parallelism of rotational sweep-based algorithms. Particularly, sDEM applies a data relocation technique before accessing the memory and computing the viewshed, thus significantly reducing the execution time. Different implementations are provided for single-core, multi-core, single-GPU, and multi-GPU platforms. We carried out two experiments to compare sDEM with (i) the most used geographic information systems (GIS) software and (ii) the state-of-the-art algorithm for solving the total viewshed problem. In the first experiment, sDEM results on average 8.8x faster than current GIS software, despite considering only a few points because of the limitations of the GIS software. In the second experiment, sDEM is 827.3x faster than the state-of-the-art algorithm considering the best case.
The use of Unmanned Aerial Vehicles (UAVs) with multiple onboard sensors has grown enormously in tasks involving terrain coverage, such as environmental and civil monitoring, disaster management, and forest fire fighting. Many of these tasks require a quick and early response, which makes maximising the land covered from the flight path an essential goal, especially when the area to be monitored is irregular, large, and includes many blind spots. In this regard, state-of-the-art total viewshed algorithms can help analyse large areas and find new paths providing all-round visibility. We designed a new heuristic called Visibility-based Path Planning (VPP) to solve the path planning problem in large areas based on a thorough visibility analysis. VPP generates flyable paths that provide high visual coverage to monitor forest regions using the onboard camera of a single UAV. For this purpose, the hidden areas of the target territory are identified and considered when generating the path. Simulation results showed that VPP covers up to 98.7% of the Montes de Malaga Natural Park and 94.5% of the Sierra de las Nieves National Park, both located in the province of Malaga (Spain). In addition, a real flight test confirmed the high visibility achieved using VPP. Our methodology and analysis can be easily applied to enhance monitoring in other large outdoor areas.
