In this paper, we present a multi-rail design as a promising approach towards achieving Terabit/s performance for advanced E-Science applications.

We evaluate the effects of system parameters such as interrupt, memory, thread and core affinities on the achievable network throughput as we scale the performance towards Terabit/s.

We demonstrate the additive effects of Thread and Interrupt affinities for achieving 20Gbps line rate throughput.

These results can help guide the development of future intelligent middleware that will automatically optimize the performance based on system parameters and conditions, thus making it easier for applications developers, who are often not systems experts, to make full use of the system capabilities.