It's been observed in several clusters in the wild that on old platforms like el6, name resolution can be slow even when /etc/hosts is used. That's because opening /etc/hosts requires the creation of an mmap, which means the kernel has to take mmap_sem for writing, and that can contend with all sorts of other system activity. Furthermore, glibc synchronizes all resolution on a single lock, so there's opportunity for contention there. See this mailing list post for more details.
We should document that for better scalability on nodes hosting many replicas, it's best to use nscd to cache this kind of name resolution too.