Parallel VSP experiment

Paul E. McGee, Robert James Ferguson

Shot record modelling of VSP surveys can be computationally time intensive. To address this issue, we parallelize the running of forward modelling code, and we discuss the performance gains that result. The performance of this parallel implementation is measured on a cluster, and we find that performance gains (see Figure 2) are achieved with only small changes to the original code.

To discuss hardware effects on performance, hardware factors are tabulated with the most expensive factors being number of processors, the processors, cluster memory and the communications backbone. We derive an equation relating total computation time to size of the survey, they are related through hardware dependent coefficients. Software implementation is most importantly done using Matlab's Parallel Computing Toolbox (PCT) and inline methods. The VSP experiment is done with computation of different shots being distributed to Matlab workers.