Archived Content

The following content is from an older version of this website, and may not display correctly.

There’s a growing body of software-defined networking advocates who say it’s time for the entire datacenter to be defined by the server.  Broadcom, for obvious reasons, is not among them.  It’s the worldwide market share leader in physical switches — some analysts calculate that its Trident line commands at or near a staggering 90%.

But last July, Cavium, one of the main brands hovering in the 10-percent space, announced it is acquiring switch silicon maker XPliant.  It was a surprisingly small deal, investment-wise, that led to a huge announcement:  Cavium will soon be integrating XPliant silicon into its next generation of switches, the CNX880xx series.  That move will put Cavium on the board with 3.2 Tbps bandwidth, with a product release sometime in calendar Q4.

Broadcom broadside
Broadcom has now answered with the announcement that the world had been anticipating, but couldn’t pin a date on:  It has begun sampling of its long-rumored Tomahawk series, matching Cavium’s numbers with a 3.2 Tbps model that’s also compliant with OpenFlow and ready for SDN.

“As network endpoints — the servers and storage themselves — increase in capacity and bandwidth, they may be reaching the scaling limits of the pipes themselves,” states Nick Kucharewski, Broadcom’s senior director of network switch product line marketing, in an interview with DatacenterDynamics.  “Servers and storage are exceeding the limits of a 10-gig Ethernet link.  Looking at the total scale of the entire network, as you get to larger and larger network sizes, the ability to grow the datacenter even larger is in many ways limited by the total bandwidth of the switches, or the number of ports on those switches, comprising the network.”

Tomahawk aims to solve that little problem by providing connectivity for up to 128 25 Gbps ports, up to 64 40/50 Gbps ports, or 32 100 Gbps ports.

Kucharewski tells us that cloud data centers are just now reaching the limits of their 10-gig pipelines.  That fact alone is already limiting the actual physical size of datacenters that major cloud providers can build.  No doubt Broadcom has already discussed this constraint with the Open Compute Project, Facebook’s initiative to develop an open specification for cloud datacenters.

The dream of the emerging field of network functions virtualization (NFV) is to absorb the entirety of network switching into software, and enable servers to directly manipulate the configurations and assemblies of those switches in real-time.  Where engineers concede ASICs have the edge over NFV is raw speed.  Although recent NFV demonstrations show vSwitches handling close to 200 Gbps bandwidth under controlled conditions, folks in the real world are struggling to crack the double-digits on that scale.




Can NFV catch up?

NFV advocates claim their architecture makes up for these deficiencies through simpler, more conservative, network topologies.  But Kucharewski counters that there’s a speed threshold, beyond which its switching fabric provides all the simplicity a network admin could ask for, by means of much fatter pipes.

“When you’re looking at network design and management, having sufficient bandwidth, and not too tightly provisioning that bandwidth, is a key to simplicity of operations, and also future scale of the network,” remarks Broadcom’s marketing director.  “That said, even with a network that has sufficient bandwidth to handle current-day and future compute needs, there is still a need for additional visibility and control to adapt to those workloads.  So I would characterize it as a two-pronged approach:  One is to make sure you have the bandwidth, and you have enough overhead within the network so that it’s provisioned properly; and then two, that you have good visibility and adaptability to the traffic flows operating on that interconnect.”

Broadcom’s value-add for Tomahawk includes a new policy-based mechanism for directing traffic flow and load balancing, that may seem familiar to any infosec professional who’s worked with programmable firewalls.  Called BroadView, this system will adopt a four-stage approach to traffic management: monitoring the live state of each active switch, analyze the packets showing up on the SDN controller, optimizing the reaction to that traffic flow, and automating the way similar reactions can be taken in the future.  It’s a way of creating “if/then” rules for traffic flow that let a software-defined network of hardware-defined components truly follow the instructions of software.

“We categorize the BroadView features into network-level analytics, and specific packet-level analytics,” explains Kucharewski.  “The intent of BroadView is for every switch within the network to feed back pieces of information to an SDN controller that can then assemble a network-wide view of congestion patterns occurring within the network.  This is a sort of ‘crowdsourced’ implementation, where each switch is contributing to a picture of the whole.”

He offers one critical example:  At every point in the network, a packet has a choice of multiple paths.  The quality of load balancing influences whether the network is flowing smoothly or experiencing congestion.  That congestion will be felt by a kind of sensor located in Tomahawk’s on-chip buffer, indicating a backlog of packets within the switch.  This accumulation takes place well before packets end up being dropped.  With load balancing monitors, an admin can see how evenly packets are being distributed, and can respond to emerging bottlenecks, which show up graphically as hotspots.  Like a weather forecaster, admins can narrow or widen the window of the timescale from a few milliseconds to several seconds.

At the packet level, he goes on, BroadView presents a picture of single points in the network, using a kind of canary-in-the-coal-mine approach.  “Then we provide packet-tracing capabilities.  For instance, for a sample packet that would be injected at that switch, how would that switch handle the packet?  This is helpful if, for example, you’re losing packets within the network.  You can debug that while the network is running.”

Broadcom has yet to provide a specific release timeframe for Tomahawk and its accompanying services, though sampling is beginning immediately.