if this feature is added to Ceph, will it be valuable to Ceph users, esp. those using Ceph for GPU training clusters?