This work shows that the seekable socket interface is useful for TCP-based data communication, but does not implement
a full-scale messaging library around this interface.
Experience with microbenchmark code suggests that few changes are required for using seekable sockets. However,
integrating the interface into an MPI library may yield further insights on the usability of this interface. For example, applications will first need to inspect message headers using traditional receives (possibly peeking) before knowing
the amount required for seeking.
The actual performance will likely vary based on the amount to which messages are received out-of-order; since our microbenchmark results only consider computation time (and not its overlap with communication latencies), these results should be composable to describe overall system performance according to the methodology of Saavedra and Smith [15]. The results should also be largely independent of the physical medium used or the number of sockets since the operating system changes are confined to the socket layer and are strictly on a per-socket basis.