Hi folks,
Mesos containerizer does not have native support for isolating UTS namespace.
Because of that, NCCL always assumes that different containers running on the same host are able to do P2P and SHM to each other: code1, code2.
Do you think it'd be reasonable to extend getHostHash() to include other isolation namespaces, e.g. /proc/self/ns/mnt?
Hi folks,
Mesos containerizer does not have native support for isolating UTS namespace.
Because of that, NCCL always assumes that different containers running on the same host are able to do P2P and SHM to each other: code1, code2.
Do you think it'd be reasonable to extend getHostHash() to include other isolation namespaces, e.g.
/proc/self/ns/mnt?