Date : June 10, 2021, 9 a.m. - EN-NEJJARY Driss - Visio-conférence
Des approches séquentielles et parallèles pour l’amélioration des performances des sélections et des agrégations des données rasters
With the emergence and the production of a large volume of spatial data, supporting large scale and high-performance queries and analysis has become crucial and essential in several applications and fields. The tremendous advances in technology such as smartphones, internet of things, web, navigation systems and sensors, have led to the production of spatial datasets having large sizes. For example, climate and precision agriculture sector are ones of the fields affected by these advances in data acquisition technology where this kind of data is produced in high precision and large temporal sequences. Querying large-scale data allows extracting more valuable and meaningful information that is vital for decision-making, scientific advancement and scenario predictions. Unfortunately, most of existing methods and approaches are based on traditional computing framework (uniprocessors) which makes them not scalable and not adequate to deal with large-scale data.
In this work, we show that using the GPGPU can reduce the time of spatial data processing and save computations. In this regard, we have proposed to speed up three classical queries that have never been tackled before in the literature. First, we have proposed an optimized parallel method based on GPGPU to produce overlapping aggregated data summaries by the computation of the average temperature for all overlapped raster subsequences of a determined length for the studied region.
As a second contribution, we have tackled a raster selection query based on a threshold fixed by the user. In fact, in different analyses, users can be interested only in some rasters. Hence, we have implemented two solutions based on the GPGPU and the CPU that include a rejection procedure of rasters in the early stages of computations using on a sorting step.
Finally, we have proposed two high-performance methods for a selection query based on GPGPU and CPU for massive spatio-temporal data. The query consists on selecting fixed size disjoint raster subsequences based on their average satisfying a user threshold condition. The two methods include a rejection procedure of subsequences based on sorting.
M. François Pinet, Directeur de recherche, Encadrant INRAE, TSCF
Mme. Myoung-Ah KANG, MCF, Encadrante ISIMA, LIMOS
M. Alain Bouju, MCF- HDR, Rapporteur, Université de La Rochelle, L3i
M. Sofian Maabout, MCF- HDR, Rapporteur, Université de Bordeaux, LaBRI
M. Olivier Teste, Professeur, Examinateur, Université Toulouse2, IRIT
M. Jean-Denis Mathias, Directeur de recherche, Examinateur, INRAE, LISC.