One Big, Complex Ethernet

Arthur Cole

It used to be that selecting an enterprise networking protocol was relatively easy. If you had the money, you went with Fibre Channel for storage, InfiniBand for the server interconnect and Ethernet for everything else. If not, you made do with just Ethernet as best you could.


These days, it appears most enterprises are willing to standardize with Ethernet as the basis for a unified fabric, which would seem to make things simpler, but it doesn't. That's because many flavors of Ethernet are hitting the channel at once, forcing networking executives to take a hard look at not only their existing network infrastructures, but the numerous types of data they carry to hopefully mix and match the right Ethernet for the job.


You can see this in the way the networking industry is starting to tailor specific solutions for specific data sets. For example, Mellanox and Arista are pitching the combination of the ConnectX-2 adapters with the 7100 10GbE switch directly at the financial sector. The low latency and high scalability factors that both devices bring to the table are ideal for the high-volume world of big money. The adapters are bolstered with the RDMA over Converged Ethernet (RACE) protocol, which trims latency to 6s at 1 million messages per second, and has the added benefit of the OpenFabrics Enterprise Distribution (OFED) API.


Ethernet can also be implemented with iWARP (internet Wide Area RDMA Protocol) if you're looking to boost performance in high-performance computing settings like gene sequencing and weather modeling. IBM and Chelsio recently ran a number of tests that showed iWARP-enabled 10 GbE deployments could outperform even InfiniBand under certain circumstances. The configuration, which consisted of a dual-port, offload-enabled 10 GbE adapter from Chelsio tied to a 20-port 10 GbE switch from Force10, did best under the NETPERF benchmark designed to gauge bulk data transfers over TCP/IP or UDP/BSI.


Ethernet is also opening up to new and varied virtual configurations now that enterprises have the means to tie virtual I/O technology directly to the network. Xsigo recently announced a new version of its I/O Director that plugs into the Ethernet ports on the server, cutting the need for converged network adapters (CNAs) and a host of other network devices. While this offers a tremendous advantage in matching the proper amount of network resources to various data loads, it places even greater demand on the Ethernet layer to be more flexible to suit a wide variety of traffic configurations at once.


And this is all coming at a time when enterprises are still struggling to make peace with the idea of multiple networking protocols all co-existing on a single Ethernet fabric. The rather unenthusiastic adoption of Fibre Channel over Ethernet (FCoE) highlights this dilemma perfectly. By this time, FCoE was expected to have widespread industry support, but according to Enterprise Storage Forum lackluster integration on the part of major storage vendors has relegated the protocol to top-of-the-rack status, which makes it harder to implement the kind of end-to-end solution needed for a truly effective Ethernet-based storage configuration.


So on one level, we have the luxury of knowing that Ethernet is likely to play the main role in enterprise networking for some time. The drawback is that there are many ways in which Ethernet can be implemented, and that means the decision-making process will be just as complicated as it ever was.



Add Comment      Leave a comment on this blog post

Post a comment

 

 

 

 


(Maximum characters: 1200). You have 1200 characters left.

 

 

Subscribe to our Newsletters

Sign up now and get the best business technology insights direct to your inbox.


 

Resource centers

Business Intelligence

Business performance information for strategic and operational decision-making

SOA

SOA uses interoperable services grouped around business processes to ease data integration

Data Warehousing

Data warehousing helps companies make sense of their operational data