we have discovered the reason for the issue: The ulimit on the _server_ side was too low and it differs between regular users and daemons like the DAOS server. For the latter it was set to soft 1024/hard 4096. We increased it to 50000 respectively by modifying the service/unit file. With that we did multiple IOR runs with up to 48 processes and SX object class from a single client node without any errors.
We noted that the coredump end memlock limits are already "increased" in the server's unit file. Maybe it is a good idea to increase the file limit as well by default, although the limit may depend on the provider in use.
