Can ucx support to deal with SIGSTOP and SIGCONT？

[razor1991]

I'm using openmpi with ucx 1.10.1, using ud tls. I want to suspend MPI with sending a SIGTSTOP signal and resume with a SIGCONT signal. But I found that when I suspend a long time, ud will timeout after I resume MPI program.

I want to ask,

1. Dose ucx support to deal with SIGTSTOP and SIGCONT for a long time with UD?
2. How about tls RC?

PS: I tried with a big value of UCX_UD_MLX5_TIMEOUT, it's OK.

[yosefe]

The above behavior makes sense. Without progress in the process, UD will consider the packet lost and may eventually time out.

[razor1991]

The above behavior makes sense. Without progress in the process, UD will consider the packet lost and may eventually time out.

Thanks for your reply. If I add a singal handler for UD in UCT to reset the timer to aviod timeout, do you think this is a reasonable design?

[yosefe]

The above behavior makes sense. Without progress in the process, UD will consider the packet lost and may eventually time out.

Thanks for your reply. If I add a singal handler for UD in UCT to reset the timer to aviod timeout, do you think this is a reasonable design?

It would work if all processes in the job are stopped and continued simultaneously - since if a process is stopped, the timeout will happen in other processes that communicate with it.