Tests using cloudbench/readdelay to determine whether the response time
to read an object depends on how long the TCP connection has been idle.
It doesn't seem that there is much of a consistent effect.