From the paper it looks like this is for read heavy workloads (testing write performance is "future work") and I think for network file systems which will add latency.