TCP-Fiber: Direct measurement optical transport congestion control for beyond 10 Gigabit networks

摘要

Optical or gigabit communication links could currently allow petabytes of data to be transferred to geographically distributed tera-scale computing facilities at beyond 10Gbps rates. While the bandwidth is available in network link technology, transport protocols like TCP/IP and common network host architectures severely limit the attainable throughput over such links. Traditional layering -that is implemented through excessive per-byte (word) memory bandwidth constrained buffer copying- transport processing complexity, combined error and congestion control and trial and error timeout-based approaches result in prohibitively increasing performance degradation as network speeds increase. In this paper we present TCP-Fiber, a TCP version that is based on direct measurements of available and bottleneck link bandwidth and is able to perform decoupled error and congestion control while supporting zero-copy from application to network interface. A key innovation in TCP-Fiber is a variable length "packet train" based method that allows sensing ultra high bandwidth related quantities in a network independent fashion with relaxed requirements to timers and system resources (as related to interrupts, system calls etc). A TCP-Fiber connection is able to fairly send at the full network rate without extensive trial-and-error convergence procedures or waiting on time-out for unacknowledged packets, while maintaining network stability.