The ability of your datacentre fabric to support millions of servers with low cost and power may not be due to advances from your major switch vendor, or to standards they may embrace.
It might be the result of an arrangement your switch vendor has with a handful of small, privately held companies developing software designed to accomplish just such a task.
One such company is Infinetics, a stealthy Massachusetts startup developing a software platform for data center switches and silicon designed to drastically lower the cost-per-gigabit/second of switch ports, and latency, as these switches support more and more servers. The company's product, which facilitates network virtualization, is intended to overcome the number of ports-per-switch limits of the fabric solutions from the major vendors as they scale to support multiple data centers in a cloud computing environment.
GIGAOM CONFERENCE: Data center startups emerging to solve virtualization and cloud problems
"We can be a threat or opportunity for architectures" from major switch vendors, like Cisco's FabricPath and Juniper's QFabric, says Infinetics Chief Marketing Officer Anthony Antonuccio.
Infinetics isn't alone in this game. Others with similar proposals include Nicira, Big Switch Networks, ConteXtream and, of course, VMware.
Infinetics is in discussion with several of the major data center infrastructure and switch vendors about adding its product to theirs. Discussions are deepest with Arista Networks, IBM, HP and Korean OpenFlow switch vendor Pica8, Antonuccio says.
Arista says the Infinetics software, and that of the other network virtualization players, will integrate with its EOS operating system to optimize cloud computing environments.
"Arista is building an open networking platform based on EOS working with several controller products to deliver 10Gbps switching that integrates with the cloud architecture to lower the cost and power while improving the utilization of cloud networking," says marketing Vice President Doug Gourlay.
Both IBM and HP declined to comment. The Infinetics software can also work with hypervisors like KVM and Citrix, company officials say.
Infinetics software sounds almost too good to be true. It's designed for "unlimited" size networks without the cost or performance "penalties" of other vendors' architectures. It's intended to provide throughput that scales nearly linearly with the number of switches and servers, while supporting all traffic patterns.
It features "long hop" topology extensions designed to simultaneously minimize latency and maximize throughput for any number of nodes. It supports a unified control and management plane to allow all switches and hypervisors in a network to appear as a single autonomically managed switch; and features scriptable security policies, QoS and load balancing commands, Infinetics officials say.
And the company claims it can reduce the cost and power consumption of fabrics constructed with Cisco Nexus, Juniper EX, Arista 7500, Brocade VDX and Force10 E-series switches by more than 13x and 6x, respectively. Therein lie both the opportunity and the threat that Antonuccio alluded to.
Though talks are progressing with Arista, IBM and HP, Cisco and Juniper are tougher nuts to crack, Infinetics officials say. That might be because they have much more skin in the data center and cloud fabric game, and want to pitch their own architectures as possessing the same attributes as the Infinetics software.
Or, maybe they just don't see a need for the Infinetics technology.
"Cisco cannot provide comment on speculation or rumors regarding unannounced product developments," a Cisco spokesperson says.
Fabric proposals based on Clos or TRILL configurations -- like FabricPath and QFabric -- are only modest topology improvements over today's data center architectures, says Infinetics Chief Scientist Chris Williams. The number of ports per switch creates an inherent scaling limit, he says.
Infinetics proposes a so-called Radix switch that can run on commodity switches as well as fabric switches from the major vendors. Radix is a mathematical term for the base, or number of unique digits a numeral system uses to represent numbers. In computers, it is the binary system or base two: zeroes and ones.
The Infinetics Radix switch creates a "true" flat Layer 2 network by suppressing broadcast discovery ARP messages and controlling MAC forwarding table sizes, Williams says. It uses proprietary signaling in an unused Ethernet frame type to optimize latency and bandwidth, implicitly support multiple paths, and support a symmetrical topology to allow rerouting around faults, adapt to changes and reconverge quickly around them, and to select alternate paths for load balancing.
For security, the Infinetics Radix switch provides a single point of authentication for all points of attachment and services at origin, and supports group-based device isolation throughout the physical and virtualized network, the company says. In so doing, the Radix switch subsumes the functionality of load balancers, NATs and firewalls, resulting in savings on network infrastructure, power and payroll, Infinetics says.
And since it uses a standard Ethernet frame type, non-Infinetics-enabled switches can forward Infinetics frames like any other traffic by just ignoring the proprietary signaling scheme, Williams says. So it can operate in a mixed environment of Infinetics-enabled and non-Infinetics switches, he says.
The Infinetics software is currently ported to Fulcrum silicon used in Arista and Pica8 switches. Many data center switching vendors are awaiting release of Broadcom's Trident+ chipset, but Infinetics officials wouldn't confirm that as the next porting platform for its software.
"The next porting will be determined by the growth in our strategic relationships," Williams says.
Beta versions of Infinetics-enabled switches are expected later this year, with commercial products anticipated for the first half of next year. The software can complement OpenFlow-enabled programmable switches -- which separates flow control from the switches themselves -- but is an alternative to data center multipathing standards like TRILL and Shortest Path Bridging, Williams says.
"We are a significant advance on TRILL and SPB," Williams says. "This is a different way on how to determine paths through the network."
Read more about data center in Network World's Data Center section.