Enabling parallelized gravitational wave inference through unified-memory computation

When inferring the astrophysical parameters of a long-duration gravitational wave (GW) signal, analyses are computationally limited by the many data points in the generated GW model and its filtering against data when evaluating its likelihood. There are many attempts at mitigating these limitations by either using a faster, but approximative, likelihood function (such as a Reduced Order Quadrature or a Heterodyne likelihood) or by parallelizing the calculations using a GPU. This second approach is in turn limited by separation of GPU and CPU calculations and the computational expense of moving data between the two processor’s separate memory banks.

Here I present an inference package introducing a new parallelization method making use of novel zero-copy processor infrastructure, where both the CPU and GPU share the same unified memory. This enables simultaneous CPU (serial) and GPU (parallel) calculations on the same data structures, which in turn completely negates the need for expensive memory transfer operations thus in practice removing the connection between computational cost and the size of the time-frequency space used.

Finally, I will show examples of previously computationally infeasible GW inference problems particularly applicable for astrophysics studies using data from next-generation GW observatories.