The evolution to software defined networking (SDN) is well underway. ZK Research (I am an employee of ZK Research) shows that almost 80% of organizations are interested in the topic, although fewer than 10% have actually deployed the technology. This means there are a huge number of organizations trying to understand the best way to deploy SDN.
One such way is to leverage the cost benefits of a white box switch with some sort of standards-based technology such as OpenFlow. Low cost hardware, industry standards and a few best practices should make for a relatively straightforward deployment.
Not so fast. Not all white boxes are created equal. While all white box switches do offer compelling economics, they are known to have some performance issues. White box switches deployed as a top of rack (ToR) need handle tens of thousands of flows.
Deployed in the aggregation and core require hundreds of thousands of flows. However, the ternary content-addressible memories (TCAM) that are typically found in these products can only support somewhere in the range of 1,000 to 2,000 flows because of the limitations of the silicon. TCAMs by themselves are not broken, the size we use today however was meant for high-touch features like ACLs. And to increase the TCAM size, means the price per port gets out of whack with what end users expect from a ToR switch. The underpowered switches also create the perception that OpenFlow itself does not scale, which is obviously going to be the case if the hardware is underpowered.
Earlier this month, white box specialist Pica8 announced it has enhanced its product to overcome these limitations. The company has a highly flexible architecture allowing it to leverage switch ASICs from different manufacturers. This lets Pica8 evolve their product as customer requirement change. Pica8 is using new switch ASICs that effectively uncap the flow count of the switch by enabling the TCAM and FIB (forwarding table) to act as a single, unified table.
Pica8 is using Cavium’s XPliant switch silicon to add Table Type Patterns (TTP) to its operating system, PicOS. XPliant is high-speed silicon that gets its differentiation from its “XPliant Packet Architecture” giving it a total throughput of 3.2TBs. The Cavium chip won the SDN Idol competition, which I was a judge for, at the 2015 Open Network Summit event.
The use of TTP enables Pica8 to scale to whopping 2 million flows. Pica8 also has switch that uses the popular Broadcom Tomahawk silicon that scales to 256,000 flows, more than enough for most environments. Why would anyone need 2 millions flows today? Well, most wouldn’t but as SDNs become more widely deployed and spine-leaf becomes the norm, there will be an increase in the number of flows. For a couple of use cases that are relevant today, Pica8 posted this blog last week.
Pica8’s architecture is built on something the company calls “vASIC” which can be thought of as “software defined silicon”. The technology abstracts the ASIC functions and handles things like memory allocation, timing between functions and provides drivers for the specific ASICs. VASIC creates the highly flexible platform to bring choice and scale to data center architects and application developers. This removes any kind of concern regarding the performance of one switch versus another based on which ASICs are being used.
If white boxes are to achieve their potential, they need to perform and scale better than they have in the past. White box vendors can create differentiation for themselves through software enhancements that optimize performance, like Pica8 has done with this release of the product.