A Congestion Control Algorithm for Data Center Area Communications

Ethernet is extending its applications to data-center area, i.e. very short distance, communications such as CPU-CPU and CPU-I/O interconnections, where short end-to-end (MAC-to-MAC) data transfer delay is the key performance factor. In this paper, we propose an extended Ethernet MAC mechanism providing end-to-end reliable and congestion controlled packet transfer that minimizes data transfer delay. The proposed MAC mechanism employs a packet retransmission mechanism and a delay-based congestion control algorithm with inline initial bandwidth probing. Since inline bandwidth probing with small number of data packets tends to overestimate the initial bandwidth, we introduced two extended mechanisms; two-stage probing and RTT prediction. The RTT prediction mechanism detects congestion several RTTs earlier than using measured RTT and thus minimize queuing delay. The two-stage probing mechanism adaptively starts up data transfers without overflowing the link. The simulation results show that the proposed mechanism achieves roughly 10 times shorter transfer delay compared to TCP and 2-3 times shorter delay compared to existing delay-based congestion controls.