Banca de DEFESA: ITALO AUGUSTO SOUZA DE ASSIS

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
DISCENTE : ITALO AUGUSTO SOUZA DE ASSIS
DATA : 14/10/2019
HORA: 14:00
LOCAL: Auditório nPITI
TÍTULO:

Intra-node and Inter-node load balancing and other scalable approaches for high-performance seismic processing

PALAVRAS-CHAVES:

Multi-scale Waveform Inversion (MFWI), Coupled Local Minimizers (CLM), Efficiency, Scalability, Auto-tuning, Coupled Simulated Annealing (CSA), Reverse time migration (RTM), Load Balancing, Cyclic Token-Based Work-Stealing (CTWS), One-Sided Communication, Distributed Memory, Shared Memory, Visco-Acoustic Modeling.

PÁGINAS: 100
GRANDE ÁREA: Ciências Exatas e da Terra
ÁREA: Ciência da Computação
RESUMO:

Seismic modeling, reverse time migration (RTM), and multi-scale waveform inversion (MFWI) are three of the most important techniques in seismic surveying. Seismic modeling simulates the wave propagation, RTM generates an image of the subsurface, and MFWI produces a wave propagation velocity model. These methods demand intensive computational cost due to a large amount of data they process and the complexity of their algorithms. Because of that, they are only implemented for parallel systems in practical. Although there are efficient parallel implementations of modeling, RTM, and MFWI in the literature, further improvement can be achieved by better exploring the parallelism in these methods and the characteristics of the current parallel systems. This research proposes coupled multi-scale waveform inversion (CMFWI), an alternative method to MFWI, which improves parallel scalability by reducing the parallel dependency between the processing of different frequency content of the data. An implementation of CMFWI using the coupled local minimizers method (CLM) is presented. L2-norm results showed that CMFWI had an inferior performance when compared to MFWI. These experiments indicate that further research is necessary to implement CMFWI as it compares data with different frequency contents. This work also introduces an auto-tuning strategy for properly choosing the optimal chunk size that reduces the runtime of a 3D RTM algorithm in shared memory systems. A coupled simulated annealing method (CSA) is employed to adjust the chunk size of work that parallel loops assign dynamically to worker threads. Experiments show that the proposed method is consistently better than two default OpenMP loop schedulers being up to 44% faster. This thesis also introduces the cyclic token-based work-stealing (CTWS) for distributed memory systems. The novel cyclic token approach reduces the number of failed steals, avoids communication overhead, and simplifies the victim selection and the termination strategy. Results obtained by applying the proposed technique to balance the workload of a 3D RTM present a factor of 14.1% speedup and reductions of the load imbalance of 78.4% when compared to the conventional static distribution. Finally, an implementation of a 2D visco-acoustic modeling is presented.

MEMBROS DA BANCA:
Externo ao Programa - 2492756 - JOAO MEDEIROS DE ARAUJO
Externo à Instituição - JORGE DANTAS DE MELO
Interno - 1543191 - LUIZ FELIPE DE QUEIROZ SILVEIRA
Externo à Instituição - REYNAM DA CRUZ PESTANA - UFBA
Presidente - 1673543 - SAMUEL XAVIER DE SOUZA