Seismic imaging on novel computer architectures

This paper describes performance results obtained by porting production seismic imaging software to a set of novel processor architectures, with particular emphasis on a cluster of GPUs. Introduction Seismic processing algorithms are constantly evolving, in close connection with innovations in seismic data acquisition technology. In addition to that, an oil company has to adequate its proprietary algorithms to easily interrelate with a small set of commercial processing packages. In order to cope with this dynamic environment, production codes for seismic imaging developed at Tecnologia Geofisica of Petrobras are designed to accommodate multipurpose algorithms under a software structure that emphasizes portability and modularity. But the lifetime of a software design may be limited by the lifetime of the computer architecture it was designed for. New trends in processor architecture (multi-core and many-core) and the ever-increasing number of processors on a cluster may trigger software redesign. The flexibility of Petrobras seismic imaging software makes it a natural candidate to test the adequacy of a set of different algorithms and a previously designed software structure to the recently arrived computer architectures, since testing demands a limited amount of change in the actual production code. This paper describes the structure of seismic imaging software developed at Petrobras, its use in daily production runs and the porting and achieved speed of one algorithm – the Kirchhoff time migration – to novel computer architectures, including a cluster of GPUs.