Sporadic "resource already mapped" cuda IPC

I am using CUDA IPC to work with memory between multiple applications.

It works well if each application is started on a separate terminal. But, if I use a ROS type callback where I open the IPC handle for each subscriber, it sporadically complains about “resource already mapped” when trying to open the IPC handle.

I do not understand what the error means. If it is already mapped, shouldn’t the open IPC handle just pass that resource pointer? I also do not know why it is sporadic, it works again later without any issues.

I cannot find this error in the IPC documentation as well. The return types for cuIpcOpenMemHandle does not list CUDA_ERROR_ALREADY_MAPPED as a return option as well.

1 Like

I have this issue in triton server
when calling RegisterCUDASharedMemory
any comments???