Error Reported When the Fork Method Is Used to Create an MPI Thread

Symptom

When MPI runs test cases such as osu_latency_mp on certain servers, a segment error is reported, but the running and functions remain unaffected. Error message example:

Caught signal 11 (Segmentation fault: address not mapped to object at address 0x3dbaf590)

Possible Cause

According to the Open MPI official document, using fork(), system(), and popen() in MPI processes is insecure. This is because these operations depend on various elements, including the OS, computer hardware, and network stack. Therefore, developers should avoid such insecure operations in MPI applications.

Workaround

You can disable the DC protocol on certain NICs to avoid this issue.

-x UCX_TLS=^dc