Performance and Portability of the GENE code on Exascale Architectures*

The GENE gyrokinetic microturbulence code is one of the constituent codes of the WDMApp (Whole Device Modeling Application) ECP project,

designated to simulate gyrokinetic microturbulence in the core of a fusion device.

As part of the exascale computing project (ECP), we have enabled GENE to efficiently use GPUs. The core of our performance portable approach is the gtensor library, which uses C++ expression templates to transparently generate GPU code targeted to Nvidia, AMD and Intel GPU architectures, supplemented by a portability layer for vendor math libraries like FFTs, BLAS and linear solvers.

We will describe our GPU porting work, and present performance results for GENE running on current Nvidia, AMD and Intel GPUs. We will also discuss parallel scalability results and limitations inherent to a gyrokinetic code, where the main computational cost lies in time-integrating the 5d+species distribution function, but given the integro-differential nature of the underlying equations, lower-dimensional moments need to be computed to solve for electromagnetic potentials, a process that also involves gyro-averaging.