Skip to main content

Resultados relacionados

Resultados totales (Incluyendo duplicados): 5
Encontrada(s) 1 página(s)

Dipòsit Digital de Documents de la UAB
Publication Capítulo o parte de un libro (BookPart)

Building representative and balanced datasets of OpenMP parallel regions

  • Alcaraz, Jordi|||0000-0002-9640-6763
  • Sleder, Steven
  • Tehranijamsaz, Ali
  • Sikora, Anna|||0000-0003-0090-4109
  • Jannesari, Ali|||0000-0001-8672-5317
  • Sorribes Gomis, Joan|||0000-0002-8868-7382
  • César Galobardes, Eduardo|||0000-0002-9729-8557
Incorporating machine learning into automatic performance analysis and tuning tools is a promising path to tackle the increasing heterogeneity of current HPC applications. However, this introduces the need for generating balanced and representative datasets of parallel applications' executions. This work proposes a methodology for building datasets of OpenMP parallel code regions patterns. It allows for determining whether a given code region covers a unique part of the pattern input space not covered by the patterns already included in the dataset. The proposed methodology uses hardware performance counters to represent the execution of the region, which is referred to as the region signature for a given number of cores. Then, a complete representation of the region is built by joining the signatures for every different thread configuration in the system. Next, correlation analysis is performed between this representation and the representation of all the patterns already in the training set. Finally, if this correlation is below a given threshold, the region is considered to cover a unique part of the pattern input space and is subsequently added to the dataset. For validating this methodology, an example dataset, obtained from well known benchmarks, has been used to train a carefully designed neural network model to demonstrate that it is able to classify different patterns of OpenMP parallel regions.
Dipòsit Digital de Documents de la UAB
Publication Capítulo o parte de un libro (BookPart)

Wind field parallelization based on Python multiprocessing to reduce forest fire propagation prediction uncertainty

  • Sanjuan Gómez, Gemma|||0000-0003-1406-3320
  • Margalef, Tomàs|||0000-0001-6384-7389
  • Cortés Fité, Ana|||0000-0003-1697-1293
Forest fires provoke significant loses from the ecological, social and economical point of view. Furthermore, the climate emergency will also increase the occurrence of such disasters. In this context, forest fire propagation prediction is a key tool to fight against these natural hazards efficiently and mitigate the damages. However, forest fire spread simulators require a set of input parameters that, in many cases, cannot be measured and must be estimated indirectly introducing uncertainty in forest fire propagation predictions. One of such parameters is the wind. It is possible to measure wind using meteorological stations and it is also possible to predict wind using meteorological models such as WRF. However, wind components are highly affected by the terrain topography introducing a large degree of uncertainty in forest fire spread predictions. Therefore, it is necessary to introduce wind field models that estimate wind speed and direction at very high resolution to reduce such uncertainty. Such models are time consuming models that are usually executed under strict time constrains. So, it is critical to minimize the execution time, taking into account the fact that in many cases it is not possible to execute the model on a supercomputer, but must be executed on commodity hardware available on the field or at control centers. This work introduces a new parallelization approach for wind field calculation based on Python multiprocessing to accelerate wind field evaluation. The results show that the new approach reduces execution time using a single personal computer.
Dipòsit Digital de Documents de la UAB
Publication Capítulo o parte de un libro (BookPart)

Early adaptive evaluation scheme for data-driven calibration in forest fire spread prediction

  • Fraga, Edigley|||0000-0002-5525-6164
  • Cortés Fité, Ana|||0000-0003-1697-1293
  • Cencerrado Barraqué, Andrés|||0000-0002-9143-3279
  • Hernández Budé, Porfidio|||0000-0002-8592-934X
  • Margalef, Tomàs|||0000-0001-6384-7389
Forest fires severally affect many ecosystems every year, leading to large environmental damages, casualties and economic losses. Established and emerging technologies are used to help wildfire analysts determine fire behavior and spread aiming at a more accurate prediction results and efficient use of resources in fire fighting. Natural hazards simulations need to deal with data input uncertainty and their impact on prediction results, usually resorting to compute-intensive calibration techniques. In this paper, we propose a new evaluation technique capable of reducing the overall calibration time by 60% when compared to the current data-driven approaches. This is achieved by means of the proposed adaptive evaluation technique based on a periodic monitoring of the fire spread prediction error.
Dipòsit Digital de Documents de la UAB
Publication Capítulo o parte de un libro (BookPart)

Cloud-Based Urgent Computing for Forest Fire Spread Prediction under Data Uncertainties

  • Fraga, Edigley|||0000-0002-5525-6164
  • Cortés Fité, Ana|||0000-0003-1697-1293
  • Margalef, Tomàs|||0000-0001-6384-7389
  • Hernández Budé, Porfidio|||0000-0002-8592-934X
Forest fires severely affect many ecosystems every year, leading to large environmental damages, casualties, and economic losses. Emerging and established technologies are used to help wildfire analysts determine fire behavior and spread, aiming at more accurate prediction results and efficient use of resources in fire fighting. We propose a novel forest fire spread prediction platform based on a proven two-stage prediction model devised to deal with input data uncertainties. The model is able to calibrate the unknown parameters based on the real observed data using an iterative process. Since this calibration is compute-intensive and due to the unpredictability of urgent computing needs, we exploit an elastic and scalable cloud-based solution platform implemented through coarse-grain parallel processing using a work queue.
Dipòsit Digital de Documents de la UAB
Publication Capítulo o parte de un libro (BookPart)

Efficient Cloud-Based Calibration of Input Data for Forest Fire Spread Prediction

  • Fraga, Edigley|||0000-0002-5525-6164
  • Cortés Fité, Ana|||0000-0003-1697-1293
  • Margalef, Tomàs|||0000-0001-6384-7389
  • Hernández Budé, Porfidio|||0000-0002-8592-934X
Every year, forest fires cause damages to biodiversity, atmosphere, and economy. To face such permanent threat, wildfire analysts rely on emerging and well established technologies to determine fire behavior and propagation patterns. Nevertheless, input data describing fire scenarios are subject to high levels of uncertainty that represent a serious challenge for the correctness of the prediction. The unknown parameters need to be adjusted, and an input data calibration phase is carried over following a genetic algorithm strategy. The calibrated input is then pipelined into the actual prediction phase. In addition, this two-stage prediction scheme is leveraged by the cloud computing, which enables high level of parallelism on demand, almost real-time elasticity and unlimited scalability. All of them at a low-cost strategy. In this paper, to obtain more accurate prediction results and efficient use of cloud resources in the compute-intensive calibration phase, we propose a new fitness function in tandem with a strict deadline policy that decreases overall processing time. In consonance with the hard-deadline-driven nature of fire extinction activities, the proposed strategies improve the genetic algorithm convergence and decrease the response time for the calibration stage, setting up an important upper bound limit to the critical compute-intensive adjustment phase. For the case study evaluated, only 3.87% of of accuracy loss is given out in exchange of a guarantee that the calibration phase will never last more than 50 minutes in the worst case.