I don't think core affinity would have helped in this situation. They had one instance that was being shared across all cores, necessitating the reference count

I don't think core affinity would have helped in this situation. They had one instance that was being shared across all cores, necessitating the reference count to be sync'd between sockets at L3 cache level. They may have seen some benefit if you had, say, 2 instances, one per CPU socket, and dedicated an instance per socket to avoid L3 cache interactions. But, in this case, it is far simpler, cleaner and better performing to not share the instance and avoid the reference counting entirely.