Home > TERATEC FORUM > Workshop 2

TERATEC Forum 2015
Workshop 2 - Wednesday, June 24 from 9:00 to 12:30
Big Data : Optimiser la prise de décision grâce aux Data Analytics

Big Data and the evolving many-core landscape: Are traditional low level programming models really enough?
Gerard GORMAN, IMPERIAL COLLEGE LONDON

Download the presentation

Seismic imaging for energy exploration is arguably the largest big data and high performance computing application in the private sector. Data from a single survey can be many terabytes in size – and advancements in instrumentation is driving a growth in data acquisition that exceeds Moore’s Law. The most accurate approach to this data inversion problem is known as Full Waveform Inversion (FWI). FWI aims to use as much of the information in the seismogram as is possible using the most complete seismic wave propagation model as is affordable for given computing resources and time-to-solution envelop. This requires innovations on many fronts: inversion algorithms; novel numerical discretisation of the governing equations (high order methods, finite difference, finite element, spectral element methods); the fullness of the physics model (acoustic vs elastic, isotropic vs anisotropic wave equations); optimisation of the implementation for modern low power many-core processors and heterogeneous computer architectures; and this might involve rewriting the software completely using a new programming language or programming model (e.g. CUDA, OpenCL).

These challenges stack up and bring the curse of dimensionality to innovation – advancements on any of these fronts can consume whole research teams and potentially involve modifying 100k’s to millions of lines of code. In our research we have abandoned this flat low level development approach. High level languages are combined with multiple layers of abstraction and domain specific languages to maintain the flexibility and expressiveness of a high level language like Python, while at the same time exploiting code generation and compiler technologies such as autotuning to generate high performance native code for different computer architectures. The result is a separation of concerns where new numerical approaches are readily evaluated while being able to generate code that outperforms hand tuned code.

 

 

© Ter@tec - All rights reserved - Lawful mention