RapidIO Says...
RapidIO Releases Flow Control Extensions for
Data Plane Applications
Enables Immediate Development of Robust, High-Utilization
Communication Systems
Capping an 18-month effort, the RapidIO Trade Association released its Flow Control Extensions Specification that provides congestion control for medium-rate data plane applications using the RapidIO interconnect architecture. These end-to-end flow control extensions enable the immediate development of cost-effective, RapidIO-based communications applications such as media gateways, radio network controllers (RNCs) and routers used in mobile networks, and complement its existing link-based flow control technology.
"Simulation tests verify that these extensions provide the high data flow and throughput needed for robust, high-utilization communications systems," says Ericsson's Lars-Göran Petersen, expert in packet switching technologies, and chair of the RapidIO Flow Control Task Group that began its efforts in January, 2002. "OEMs will benefit from the wide range of standards-based, interoperable products ranging from RapidIO silicon to board and box-level solutions."
"Agere is excited about the release of these flow control extensions because it's the best step forward in the evolution of RapidIO," said Mike Elser, senior marketing manager with Agere Systems. "Flow control extensions add to the number of applications within communications equipment that can be implemented using the RapidIO interface with either ASICs or standard products. Most importantly, hot pluggability gets more robust and more implementable with flow control." Part of Agere's contributions to the work group was extensive simulation of flow control in a RapidIO system.
Ronald Luijten, manager of the server interconnect fabric group at IBM's Zurich Research Laboratory, adds: "This new extension increases the performance of the RapidIO standard by avoiding congestion at high utilization. The significant progress is based on a congestion control scheme we have contributed to RapidIO leveraging our unique expertise in interconnect networks."
Jag Bolaria, senior market analyst with The Linley Group, notes, "The RapidIO trade association is building a common set of OEM requirements for data plane applications. For access, WAN, and MAN systems, these flow control extensions are needed for efficient bandwidth utilization across the data plane fabric. A single set of OEM requirements benefits semiconductor vendors and enables multiple suppliers for OEMs. These flow extensions should help OEMs and semiconductor vendors to focus resources on developing carrier-grade solutions."
Sam Fuller, president of the RapidIO Trade Association, said: "These extensions demonstrate how RapidIO is a vibrant specification driven by its members who truly represent the embedded industry, versus a one-company standard. That's why the momentum behind RapidIO continues to increase."
These approved flow control extensions -- ready for immediate design-in -- can be downloaded from the RapidIO website at www.rapidio.org. The RapidIO Trade Association also has created a task group that is working on specifications which further expand the RapidIO marketplace by addressing the more demanding technical requirements found in carrier-grade telecommunications data plane applications.
Also available at the website is information on system-enablement tools including RapidIO vendor product lists, synthesizable Verilog cores, analog physical layer cores, logic and protocol analyzers, operating system support, bus functional models, and hardware interoperability platforms.
About the RapidIO Trade Association
The RapidIO Trade Association was formed in June 2000 to drive the adoption of open-standard, high-performance interconnect architectures needed for high-performance networking, communications and embedded systems. With more than 50 members worldwide, this non-profit organization is headquartered in Austin, Texas. Membership provides early access to the specifications, the ability to propose changes to the RapidIO standards, and the opportunity to actively participate in the adoption process. A complete list of member companies, as well as education and design tools, are available at the association's website.
analogZONE Says…
I've always had a soft spot in my heart for RapidIO (RIO) -- mostly because it's a more democratic, grass-roots interconnect standard than some of the others who are vying to dominate innards of the next generation of communications equipment. With a processor-agnostic architecture, it seems to really have been intended to support and grow the industry as a whole rather than promote a single party's interests or maintain dominance of a particular chip or box maker.
The release of their new flow control spec is particularly good news for me because it will allow RIO to move beyond its original mission of providing intelligent from chip-to-chip connections, and enable it to support board-to-board and system interconnect functions as well. It also opens up the door for medium-rate (sub-10 Gbit/s) dataplane applications.
RIO started as a technology-agnostic next-gen processor interconnect, mostly for control plane applications, but has evolved into a more general-purpose means of data transport as several manufacturers decided to migrate the technology to absorb or replace their proprietary mechanisms.
When RIO was solely for control plane, its focus was on performing relatively simple load/store and messaging functions with low latency and high transfer rates. But now, as RIO moves to the data plane, its applications also require support for things like message passing (protocol encapsulation), support for different traffic classes, and dynamically changing traffic patterns and priorities. Data plane flows are also subject to latency and QoS requirements.
You need complex mechanisms like this because once you are in a large system with lots of line cards feeding a switch fabric, you get phenomena like higher-order head-of-line blocking that can significantly degrade service when the system is heavily loaded. Higher-order HOL blocking occurs mostly in a multi-stage fabrics where a congested switch node in the fabric ends up back-propagating congestion to prior nodes and creating a non-linear decline in service that locks up part of the switch. A carrier-class system must also handle single and multiple point failures gracefully and manage the remaining traffic fairly without additional blockage -- something the original RIO spec did not address.
In order to apply RIO's efficient architecture to connecting systems rather than just processors, the folks designing the spec had to addresses concerns for congestion management and control found in more complex systems. Without those mechanisms in place to take care of 2nd order effects, rare instances and unusual circumstances can cause log jams that can cause a system's performance to plummet. To make sure that this does not happen, the RIO flow control spec stack compliments the other logical specs (I/O system, message passing Global shared memory, etc.) above it, and the common transport spec below it.
The new RIO switch spec adds an intelligent congestion management algorithm to detect, manage, and relieve congestion without losing the simplicity and efficiency of the original spec. To do this, it relies on an in-band backpressure mechanism that signals the upstream node to throttle back using a new logical layer packet type that carries "XON-XOFF" messages between switch nodes. This end-to-end flow control adds to the link-to-link control mechanism that was originally built into the RIO architecture. It provides a source-aware mechanism that directs back-pressure to the source node, rather than throttling the intermediate nodes.
The packet that carries the signaling messages has the highest priority, allowing it to pass through network without obstruction -- even during heavy congestion. For the moment, the mechanism is a hard on/off control, but it has "hooks" that will enable the standards committee to add variable rates if needed later. The real magic is in the algorithms that keep the system from dithering, oscillating, or otherwise throwing it into fits. The upstream node can get multiple XOFFs, from downstream nodes, allowing it to throttle traffic on a per-stream basis.
Since it's intended for telecom and other high-rel apps, the flow control mechanism must tolerate lost XOFF or XON messages. At least in theory, the standard RIO retransmit protocol takes care of this with only a minor performance degradation.
The result of introducing the flow control mechanism is that you can run a network much closer to its total line rate -- typically 80% - 90%, much higher than with a normal switch architecture. Carriers and network operators will appreciate this because it enables them to get much more out of their equipment before having to buy more capacity. I expect this will allow RIO to find a home as the interconnect medium in switch fabrics for DSLAMS and wireless data switches. Follow-on work should make it suitable for even larger, more complex applications deep inside carrier backbones.
|