StreamProcessor overview From Neo Networks. Rethinking Switching Technology By IDC. Insight Report on NEO Networks Technology By Sterling Research. Cree is vice president of market-ing at Neo Networks, a maker of massively parallel routers. He can be reached at mark@ neonetworks.com.)

By Mark Cree
Network World, 3/30/98

Network forwarding devices designed with dedicated silicon-based routing and switching ASICs will likely fail to keep pace with the growing performance and processing needs of today's networks.

A new technology promises to help drive future backbones: massively parallel routers.

Massively parallel routing systems in the backbone can process packets or cells in real time. That lets the routers establish network quality of service, dramatically increasing network intelligence and boosting router performance.

The fundamental problem with existing router/switch designs is that they continue to follow a development cycle that worked well when network traffic growth was measured in double digits.

The traditional approach isto find or build the biggest bus -- and more recently, switch fabric -- possible and then port existing software into accelerated hardware. With the Internet and corporate intranet installations growing by as much as 1,000 percent per year, traffic generated will quickly outpace products developed with traditional designs and dramatically reduce their useful lives.

What's needed to move the industry forward is a new approach that focuses resources on the real problem: processing power.

Routing, bridging and switching are really only different levels of data processing. Complete routing requires a lot of CPU cycles with table lookups and other header processing, while bridging and switching require less.

An architecture able to manipulate and process data at multigigabit rates easily can forward at the same rate. That's where the massively parallel approach to routing comes into play.

Parallel processing routers are different from generic gigabit router and switch products because they view incoming data as a stream of packets instead of individual packets. With parallel processing, the incoming packet or cell becomes an instruction set for the routers parallel CPU.

Massively parallel systems solve processing and data manipulation issues by distributing network intelligence over several devices or CPUs.

While one processor handles application and protocol processing for the datastream, another performs lookup, queuing and prioritization for the datastream. With this approach, routing, switching and bridging decisions become rules applied to the data stream by the processors.

Vendors build parallel routing systems in two ways. Some use a mesh of CPUs in one device to process frames in parallel (Neo Networks Inc.). Others mesh complete devices for scaling total system throughput (Pluris Inc. and Neo Networks). Although each approach is different, both move processing power close to the bottlenecks and distribution tasks.

Window of Time On a 1G-bit-per-second connection, there is only about 0.6 msec to operate on an incoming 64-byte packet. In that time window, very few operations beyond bridge/route table lookup can happen when processing in a serial mode.

Fast table lookup performance is critical in any system and quickly consumes the majority of the time window unless you perform operations in parallel.

Parallel processing systems perform multiple operations in parallel during the 0.6 msec time window. Neo Networks' approach embeds CPUs into Application Specific Integrated Circuits that run routing/switching software. As a packet is received, it is parsed and its components are sent to different CPUs for processing in parallel during the same time window.

Route lookup occurs at the same time as packet characterization for security, traffic shaping, real-time monitoring and the enforcement of other network policies that may operate on data from any part of the Open Systems Interconnection model.

Parallel systems can do the following:

• Support advanced network services such as quality of service policies
• Support high-performance backbones
• Provide longer product life cycles because code is not fixed into silicon

The Bottom Line The net benefit of processing data in parallel is that the device makes data characterization decisions within the same time span in which Layer 2 and 3 forwarding decisions are made.

Advanced data characterization facilitates establishing networkwide policies and leads to more intelligent networks. Future routing/switching architectures must advance network intelligence and performance.

Data forwarding and characterization are separate operations that will need to be per-formed at multigigabit and terabit rates to satisfy the needs of today's Internet technology-based networks.