The issue is caused by the libhwloc
library, which is used by OpenRTE
(the OpenMPI component) to explore the hardware topology of the node during the launch of parallel processes. When hwloc_topology_load()
is called, libhwloc
iterates over all of its available components, which are implemented as plug-ins. One of these components, gl
, is responsible for gathering information about the system's graphics devices and uses X11 calls like XOpenDisplay()
, which can generate the error message "Invalid MIT-MAGIC-COOKIE-1 key" if it fails to connect to the X11 server.
In many cases, this gl component is not needed on HPC clusters and can be safely disabled by setting the HWLOC_COMPONENTS
environment variable to exclude it:
export HWLOC_COMPONENTS="-gl"