Startup Neterion Inc. has rolled out a chip that lets systems split one 10-Gbit Ethernet link into as many as 16 virtual connections. The effort marks the latest milestone in a long quest to deliver a Swiss Army knife of networking capabilities for the data center in a single slice of silicon.

Virtualization is seen as the next hot feature for the servers that drive the Internet and back-end business computing. Ultimately, server makers hope to meld the technology with many others in chips that handle a combination of networking, storage and clustering jobs.

That goal is probably still more than a year away. In part, that's because the feature requirements list keeps growing, forcing respins of the chips along with sometimes-Herculean software-development efforts for devices that serve what is today a tiny market, albeit one with huge potential.

Neterion (Cupertino, Calif.) is one of a handful of companies competing to be the first to deliver all those features on advanced forms of Ethernet. The group faces competition from Mellanox Technologies, which in the final stages of testing and writing software for 40-Gbit/second Infiniband switch and adapter chips. Versions that will handle Ethernet and Fibre Channel are expected to ship in the second half.

Neterion's X3100 series chips carve parallel, bidirectional paths in silicon to let 16 independent jobs use a single 10-Gbit Ethernet link simultaneously. It implements the new single-root I/O virtualization (SR-IOV) standards from the PCI Special Interest Group and proprietary techniques found in the latest software from virtualization specialist VMware.

"This is a cost and a performance breakthrough, because the dirty little secret of virtualization is that as you start piling apps on a server, you can bring it to its knees," said Neterion CEO Dave Zabrowski.

The startup says its chip delivers throughput of 16 Gbits/second across its two 10G links, even under heavy virtualization loads, compared with 2 to 4 Gbits/s for some competing chips. The device, which dissipates 12 watts on average, is tailored for servers using as many as four quad-core processors. It will be in production in April at an undisclosed cost.

Neterion recently became the first chip maker to support the NetQueue techniques VMware baked into its ESX version 3.5, released in December. The chip also moves from PCI-X to the 2.5-GHz version of PCI Express, which will be used widely in next-generation servers.

The startup has gotten an edge in virtualization in part because of its longstanding work with IBM and Hew- lett-Packard on proprietary Unix systems. "They are ahead of everyone else in virtualization, and getting there early, they could become a de facto standard," said Bob Wheeler, an analyst with The Linley Group (Mountain View, Calif.).

The VMware software could be an early driver for 10G Ethernet deployments, Wheeler said, since its queuing technique helps 10-Gbit devices achieve something closer to their full throughput when running the virtualization software that OEMs and end users want so as to get the most bang for the buck from servers.

"Virtualization has the potential to drive 10G Ethernet adoption, but until now the products haven't had good enough performance," Wheeler said. With the ESX 3.5 software, "now the technology becomes attractive for high-end servers," he added.

Designers of virtualization software say they will support the new PCI hardware standard, but in some respects it represents a step backward. "Hardware vendors tend to spin this as a panacea, but it is not," said Simon Crosby, founder of Xen Source, an open-source virtualization-software supplier now part of Citrix Systems, where Crosby is a group chief technologist.

The PCI SR-IOV spec "breaks the model of having guest software that runs independently of hardware, and makes the software need to think about hardware dependencies again," Crosby said. "In general, we see significant improvement [with PCI IOV] in workloads where you are really hammering the system with virtualization."

But developers are still working through concerns about reliability with sampling hardware and beta code, said Crosby. "This won't be in volume hardware until sometime next year, and it is not clear exactly what it means to end users yet," he said

The trouble is that chip makers implement the PCI spec in unique ways, forcing software to follow the particularities in the chips' software drivers. Problems arise when a virtual session on one chip fails and has to shift to a session on another chip that may have a slightly different feature set.

Pieces of the puzzle
Virtualization is just the latest in a laundry list of advanced networking features designers are trying to pack into Ethernet in hopes of creating a unified data center fabric for servers, switches and storage arrays.

Like Neterion, startups NetXen and ServerEngines are prepping 10G silicon supporting the latest PCI Express links and virtualization standards. Each de- vice has its own mix of features that capture a snapshot in time of where the industry is driving the technology.

Thus far, no one has gotten all the pieces of the puzzle together yet," said analyst Wheeler.

For example, no chip today delivers both full support for virtualization and TCP offload engines (TOEs), which are used to reduce overhead on host CPUs, Wheeler said. As a crop of chips approaches that goal this year, the industry will require support for yet another emerging standard to run Fibre Channel storage traffic over Ethernet.

"I don't know if we will get to one chip with everything in it, but to flesh out this converged-fabric concept, you need to get as much capability out there as possible, and we are still early in this effort," said Michael Krause, an interconnect specialist in the PC server group at Hewlett-Packard Co. "It's a longer period of time before you get all this in the chips in a way that's cost- and power-efficient."

In the meantime, 10G Ethernet products of all sorts have been relatively expensive and slow. That translated into sales of just 50,000 such products last year, according to Linley Group estimates, penetrating a tiny fraction of a market of more than 10 million servers sold annually.

In April, ServerEngines (Sunnyvale, Calif.) hopes to sample a new version of the 10G Ethernet chip it announced in July. The device will step from 2.5- to 5-GHz PCI Express links and will support the SR-IOV specification.

The follow-on PCI Express standard, called multiroot IOV, should be com- pleted officially by April. The spec enables jobs from more than one server to share I/O resources.

The total throughput of ServerEngines' dual-ported chip will rise from about 13 Gbits/s to full-rated speeds closer to 20 Gbits/s, thanks to the 5-GHz PCI Express links. "People are dead if they don't have that," said Kim Brown, vice president of business development for the startup.

ServerEngines has drivers for the VMWare ESX 3.5 software in certification now. Whereas its current, 130-nanometer chip supports 32 separate protected domains to handle virtualization, the next-generation, 90-nm part increases that to 64 domains.

For its part, NetXen (Santa Clara, Calif.) expects to wait until fall to roll out a chip supporting 5-GHz Express. That's when Intel Corp. is expected to launch its first server chip sets supporting the link as part of its 45-nm Nehalem family of processors.

NetXen says it gets one-way throughput of about 9 Gbits/s per 10G port under the latest VMware software, although CPU utilization levels have gone up. The drain on host processors will ease when support for the SR-IOV standard kicks in toward the end of the year, said NetXen president David Pulling.

Soft challenge
NetXen implements many of the new features in firmware on four proprietary processor cores on its chip. "There's a tremendous amount of validation," said Pulling. "We are running on 1,000 servers here, and with two OEMs supporting as many as 68 variations of Linux, each with its own test matrix."

The situation is similar at Server- Engines, which has a chip using one core from ARM and eight from Tensilica. The company's first chip was completed in March 2006, but final software was not finished until last month.

"It takes that long because the amount of firmware and software is massive," said Brown. The 10-Gbit Ethernet world "is not like Gigabit Ethernet, which was just networking. This is TOE and iSCSI and virtualization."

The general-purpose processors ServerEngines and NetXen use help them stay flexible amid changing requirements. But they tend to draw more power and provide less performance than dedicated hardware based on state machines like the Neterion chips, said Neterion CEO Zabrowski.

Trafficking in storage
On the storage front, both ServerEngines and NetXen now support iSCSI, the standard for running SCSI storage traffic on Ethernet. ServerEngines got praise at its debut last year for handling the protocol-heavy iSCSI at data rates up to 8 Gbits/s.

The iSCSI standard is seeing growing adoption rates, especially at small companies. Most large companies, however, use dedicated Fibre Channel networks for storage.

"The amount of iSCSI storage capacity is large and growing, but it is mainly at the low end and typically outside the United States, which is a stronghold of Fibre Channel," said Krause of HP.

The industry kicked off a standards effort last year to run Fibre Channel over Ethernet (FCoE) as part of its drive to converge all data center traffic on Ethernet. The work has been making rapid progress, with the T11 standards group announcing a milestone recently: completion of an addressing portion of the spec. A final standard and initial chip products based on FCoE are expected as early as the end of the year.

Companies such as NetXen and ServerEngines expect to rev their software again to support FCoE. The standard can require a hardware redesign for optimal chip support of all of its features.

Zabrowski said Neterion is investing in iSCSI technology but is not supporting it in products yet. FCoE is further out and still needs to be supported better in operating systems and applications, he said.

"This is a development year for unified fabrics. It is really the year of virtualization," Zabrowski said.

Clustering is another requirement for some users who want to drive all traffic to Ethernet. That typically requires support for a TOE and the remote-direct-memory access standard to deliver low latency. Chip makers are at different stages in their plans for supporting TOE and RDMA today.

Here again, Zabrowski maintained the software is not ready. "The Linux community has rejected TOE," he said, while virtualization software does not support it and chip makers that write the code themselves are adopting "an unsustainable model."

For its part, Neterion supports a so-called large-resource offload--a subset of TOE that the startup claims delivers most of TOE's advantages.