At Virtual SC20: Fraunhofer’s Dr. Daniel Grünewald Talks GaspiLS, a Scalable Linear Solver Library for the Exascale Age

Print Friendly, PDF & Email

At Virtual SC20, Dr. Daniel Grünewald of the Fraunhofer Institute in Germany updated us on GaspiLS, a scalable linear solver library that the institute says is built for the exascale age. Many engineering simulations are based on CFD and FEM methods to, for example, determine aerodynamic properties of planes or analyze statics of buildings. This can consume much computation time, and that’s where GaspiLS comes in, Grünewald told us.

It’s based on single-sided, truly asynchronous RDMA-based communication primitives. (They) are supplemented by a lightweight synchronization mechanism. Every thread can essentially communicate inside of the model. And it allows for perfect overlap of communication and computation and reduces the synchronization between the communicating processes. And since GaspiLS is based on this communication model, it benefits from the scalability features.

Grünewald said the performance of iterative solvers significantly impact total run time of these simulations. GaspiLS is designed to gain faster insights from simulations to generate more detailed models, more precise parameter studies and more cost-efficient resource utilization.

To maximize scalability, GaspiLS uses tools developed at the institute for parallel programming, including the communication library GPI-2 and its underlying programming model. The algorithm allows for the assignment of executable tasks to free compute resources at any time and guarantees for a continuous stream of compute tasks for every CPU.