Ethernet at iWARP speed

The world of high-speed network connections has been expensive, complicated and difficult to standardize, but iWARP, a technology that has been fermenting for a couple of years, could change all that in the next 12 months, according to Ethernet experts.  

iWARP uses a concept called Remote Memory Direct Access (RDMA). It puts routing and processing functions into silicon, which lets it communicate between applications without needing to talk directly to the operating system.  Writing directly to another computer’s memory while bypassing the sending computer’s processor enables data centres to scale their networks without increasing server loads. 

“Remote direct data transfer has been around for a while. It’s a question of making it available for Ethernet,” said Info-Tech Research analyst William Terrill. This is its main advantage over InfiniBand, the incumbent high-speed interconnect protocol, which also uses RDMA. “iWARP is a good option because it’s built over TCP/IP and as such you don’t need special Infiniband switches or cable,” said Terrill. “It’ll pick up very quickly in the iSCSI world, for connecting with SANS.”

Although things have been quiet in the iWARP world for a couple of years, we can expect activity to increase soon, according to Bob Noseworthy. As technical director of the University of New Hampshire’s Dartmouth Testing Lab, Noseworthy has been testing iWARP interoperability as part of a consortium. Other members of the consortium include Hewlett Packard, NetEffect, and NetApp, all of whom have been part of the testing group since 2004. Others include Chelsio, HP. Lamprey Networks, EMC, Broadcom, Intel, and Sun. “A lot of these folks have been making inroads around ASIC designs and we’ll be hearing from them very soon,” Noseworthy said.

Standardization work currently in preparation could pave the way for iWARP. The Open Fabrics Alliance (OFA), an organization designed to promote the development and adoption of RDMA-based technologies, is currently preparing an open source communications stack for the protocol.

The software stack, designed to run as a small Linux-based executive program on iWARP boards, will provide a standard way for iWARP vendors and customers to implement the protocol. The lack of a standard stack for InfinBand was one reason for its relatively slow take-up in the past few years, Bill Boas, vice chair of the OFA. It was also the reason that the OFA (originally called OpenIB) was formed in the first place.

“As a customer who wanted to use Infiniband we didn’t want to have to figure out whether we were using this or that vendor’s stack,” said Boas, who struggled with this problem when managing Infiniband connections at Livermore Labs. “We didn’t want to buy InfiniBand hardware from just one vendor. Buying the hardware is the easiest part of it. Making the software work together is the most difficult.”

But just as things shape up for iWARP, InfiniBand is to receive a much-needed shot in the arm too, said Boas. The OFA, which also supports InfiniBand, has an open source stack to standardise that technology, too. With the InfiniBand specification soon expanding to support iSCSI capabilities, this will bring the technologies closer together still.

InfiniBand, however,  is attempting to move downwards from the high-performance computing world, whereas iWARP applies high-performance RDMA concepts to what has traditionally been a lower performance, commercial technology. 

“We view iWARP as a Band-Aid to Ethernet to get it up to the performance of InfiniBand,” argued Len Rosenthal, who is involved with the InfiniBand Trade Association. Infiniband reaches 20 Gbps today and will move to 40 Gbps soon, he says. For the time being, iWarp will top out at 10 Gbps per second and its ability to aggregate links is not clear.


Share on LinkedIn Share with Google+