Artificial Evolution : a Complex System that can efficiently exploit massively parallel computing eco-systems to solve inverse problems with EASEA and EASEA-CLOUD

Mardi 18 septembre 2012 à 11h00, Amphi 6, Bâtiment "Le Patio", 22 rue Descartes, Université de Strasbourg.

Premier séminaire mensuel du CNSC Strasbourg

• Pr Pierre Collet, Artificial Evolution : a Complex System that can efficiently exploit massively parallel computing eco-systems to solve inverse problems with EASEA
  ICUBE Laboratory, Strasbourg DCCS, Strasbourg University
  
  

Click on the following links to obtain the powerpoint presentation (divided into 3 files because of a 50MB restriction on file sizes): Media:CNSC0912a.pptx Media:CNSC0912b.pptx Media:CNSC0912c.pptx

With the advent of GPGPU cards, all computers are becoming massively parallel systems that are very difficult to program efficiently. Indeed, the last generation of NVIDIA cards (GTX690) provides 3072 cores grouped into multi-processors of 32 SIMD cores. This means that in order to efficiently execute an existing algorithm on a single PC with a hexacore CPU and a top notch GPGPU card, one should decompose the algorithm into 6 major tasks and 3072 minor ones which, 32 by 32, should execute the same instruction at the same time!

Super-computers are currently created by putting together thousands of such machines (TSUBAME 2.0, 5th machine in the Nov 2011 top 500 super-computer ranking is made of 1442 nodes containing 3 GPU cards each) but because of necessary synchronizations between cores and data exchanges, exploiting efficiently such machines (standard PCs with GPU cards, clusters of GPU PCs, GPU super-computers) is virtually impossible with standard top-down algorithms.

Fortunately, Complex Systems produce results that emerge from the multi-level interaction of many independent entities that can be directly implemented in a very efficient way on multi-level massively parallel machines or computing eco-systems (made of potentially heterogeneous computers, clusters, grids, clouds, super-computers) allowing to use exascale computing facilities in an efficient way.

This talk will show how Evolutionary Algorithms (that implement a complex system) can be ported on such computing eco-systems in order to optimize nearly any kind of continous, discrete, combinatorial, or mixed problems thanks to the EASEA-CLOUD massively parallel evolutionary computing platform.