Some programs on “maxwell” and “pascal” GPU nodes may need to be upgraded following May downtime
We are going to upgrade the underlying operating system on the GPU nodes in the next downtime two weeks from now. As we tested we found that the OpenMPI associated with GCC 5.4 (OpenMPI/1.10.3) does not work with the newly installed driver, so if the system gets upgraded then the 1.10.3 version of the MPI library (with GCC 5.4) can’t work properly. This will affect all of the associated applications building on top of this version of MPI on
pascal nodes (
turing nodes are not affected). The affected applications (version match the given MPI) are listed below:
We note that the higher version may not be affected by the upgrade. This software list may not be a complete list; in general, if the running applications are on the
pascal nodes and requesting multiple nodes, then the job is an MPI job on the GPU nodes. Please check whether the application is in this list (version is important, higher version may not be affected). If you use OpenMPI 1.0.3 and GCC 5.4 to compile the code, please also double check it. If you are affected by this upgrade please contact us through the ticket system, and we will help you to smoothly transform to a newer version.