Abstract
IPv6 is the response to the shortage of IPv4 addresses. It was defined almost twenty years ago by the IETF as a replacement of IPv4, and little by little, it is becoming more preponderant as the Internet protocol. The growth of Internet has led to the development of high performance networks. On one hand, Ethernet has evolved significantly and today it is common to find 10 Gigabit Ethernet networks in LANs. On the other hand, another approach for high performance networking is based on RDMA (Remote Direct Memory Access) which offers innovative features such as kernel bypass, zero copy, offload of splitting and assembly of messages in packets to the CAs (Channel Adapters), etc. InfiniBand is currently the most popular technology that implements RDMA. It uses verbs instead of sockets and a big effort of the community is required to port TCP/IP software to InfiniBand, to take advantage of its benefits. Meanwhile, IPoIB (IP over InfiniBand) is a protocol that has been proposed and permits the execution of socket-based applications on top of InfiniBand, without any change, at the expense of performance. In this paper, we make a performance evaluation of IPv6 and IPv4 over 10 Gigabit Ethernet and IPoIB. Our results show that 10 Gigabit Ethernet has a better throughput than IPoIB, especially for small and medium payload sizes. However, as the payload size increases, the advantage of 10 Gigabit Ethernet is reduced in comparison to IPoIB/FDR. With respect to latency, IPoIB did much better than 10 Gigabit Ethernet. Finally, our research also indicates that in a controlled environment, IPv4 has a better performance than IPv6.
Highlights
Internet Protocol version 6 (IPv6) [1][2][3][4] is around for some years
We report the throughput and latency obtained at the level of UDP and TCP, for IPv6 and IPv4, when varying the payload size
Our experiments indicate that the throughput of 10 Gigabit Ethernet (10GbE) outperforms the ones of IP over InfiniBand (IPoIB), except for very large payload sizes (e.g. 32,768 bytes) where IPoIB in full-duplex signal rate of Gbps (FDR) has the best throughput
Summary
Internet Protocol version 6 (IPv6) [1][2][3][4] is around for some years now. It is a solution to the unexpected dramatic growth of the Internet, which is facing the exhaustion of available IPv4 addresses. That is with RDMA, the CA (Channel Adapter) directly places the application’s buffer into packets on sending, and the content of the packets into the application’s buffer on reception, without any intervention of the CPU. This allows a much better communication system with zero copy. InfiniBand defines its own stack of protocols It does not use sockets as TCP/IP applications do, and is based on “verbs”. We make a performance evaluation of 10 Gigabit Ethernet and IP over InfiniBand, where the former is the new de facto technology for local area network.
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call