diff --git a/src/docs/JOSS/paper.md b/src/docs/JOSS/paper.md index 10791ff083f..6cc4d9477b2 100644 --- a/src/docs/JOSS/paper.md +++ b/src/docs/JOSS/paper.md @@ -193,7 +193,7 @@ Simulations such as this will play a critical role in predicting the viability o As an example of the weak scalability of GEOS on an exascale class system, we present two weak scaling studies on a simple wellbore geometry run on the Frontier supercomputer at Oak Ridge National Laboratory. Frontier is comprised of 9,472 Cray EX235a nodes that each contain a single AMD EPYC 7A53 CPU and four AMD MI250X GPU's [@frontier]. Note that each MI250X is comprised of two Graphics Compute Dies (GCD), with each GCD appearing as a GPU to the operating system. -A more detailed discussion and instructions to reproduce the results are available in the [Performance Benchmarks](https://geosx-geosx--3296.com.readthedocs.build/en/3296/docs/sphinx/advancedExamples/performanceBenchmarks/Index.html) of the GEOS documentation. +A more detailed discussion and instructions to reproduce the results are available in the [Performance Benchmarks](https://geosx-geosx.readthedocs-hosted.com/en/latest/docs/sphinx/advancedExamples/performanceBenchmarks/Index.html) of the GEOS documentation. The weak scaling results for mechanics are presented in (Figure \ref{fig:Frontier_scaling}a) and shows nearly flat scaling of the GEOS processes (assembly/field synchronization) up to 32,768 GPU's($81.3 \times 10^{9}$ degrees-of-freedom). There is a moderate decrease in efficiency with the application of the hypre preconditioner setup and solve, but given the complexity of those algorithms, this level of scaling efficiency is excellent.