OmTCP: Increasing Performance in Server Farms

Abstract

Normal TCP/IP operation is for the routing system to select a best path that remains stable for some time, and for TCP to adjust to the properties of this path to optimize throughput. By executing TCP’s congestion control algorithms on multiple paths at the same time, a multipath TCP can shift its traffic to a less congested path, thus maximizing both the throughput for the multipath TCP user and leaving more capacity available for other traffic on more congested paths. And when a path fails, this can be detected and worked around by multipath TCP much more quickly than by waiting for the routing system to repair the failure. This paper proposes a one-ended multipath TCP that is implemented on the sending host only, without requiring modifications on the receiving host, for the purposes of maximizing performance in transmissions from multiply connected large servers towards singly connected end-users and recovering from failures more quickly.