high performance computing on graphics processing units: hgpu.org

Graphical Processing Units (GPUs) can provide considerable computational advantages over multi-core CPU nodes or distributed networks by locally accelerating certain types of floating point operations. However, when processing and inverting exploration scale seismic datasets we encounter two key problems – compounded disk IO (explicit routing through the host is necessary) and the relatively small memory provided by the GPU (<=6 Gbytes, restricting model sizes that can be allocated). As shown in an earlier discussion the IO bottleneck on the adjoint side can be somewhat circumvented by using random domain boundaries. Herein will be discussed how the forward modelling routine must be adapted to create an adjoint pair such that least-squares iterative inversion can be performed. We will then analyse how domain decomposition and P2P communication can be used to propagate over larger model sizes in such a way that communication can be effectively hidden and subsequently we can observe linear scaling.