It would be interesting to see the client/benchmarking program. It almost sounds like it could be single-threaded ... which would mean the delay is an artifact of the benchmark only having one op outstanding, rather than something inherent in the storage layer.