re: Foundry Strikes at the Core"Apparently Foundry has built a router based on this principle."
arch_1, good primer. What I read is: data plane has limited autonomy (like the states) while control plane has overall control and will interfere when necessary (like the fed). Does that mean one needs a cpu plus an npu/asic to build the router, or only one chip can do the job? If so which one? TIA.
Router people use strange terminology that may not be obvious to hardware peippe. We speak of the "data plane" and the "control plane." The router "data plane" operates on each packet, based on information that is manipulated by the "control plane." If the per-packet decision can be made sufficiently simple, They can be done in cheap high-speed "hardware" of some sort. We frequently use the term "forwarding" to refer to the per-packet processing.
In the control plane, we run comples routing protocols. Routers use these protocols to communicate with each other to determind the current topology o fhte network, and to determine the best egress for each ingress packet. The control plane protocols are so complex that they must be implemented in software on general-purpose CPUs. To a first approxoimation, the amount of processing power needed to implement the routing protocols is independent of the interface line speed, and is only weakly coupled to the number of interfaces on a router.
Thus, "forwarding" (data plane processing) can be done in firmware. "Routing" ( control plane processing) is done in software. The results of hte routing decisions are conveyed from the contro plane to the data plane via "tables."
If the core is fundamentally simple, then forwarding requires few decisions and the high-speed hardware is cheaper. If the core is complex, then forwarding requires more decisions and the high-speed hardware is more expensive. This is the reason that the core should be kept simple. The core should do IP routing and bulk QoS,and nothing else.
Apparently Foundry has built a router based on this principle.
re: Foundry Strikes at the CoreThe routers don't need to be OSMINE compliant and never were. OSMINE compliance is necessary of transmission and transport equipment (that too only for ILECs/RBOCs in NA). Not for switches or routers.
re: Foundry Strikes at the CoreIf the core is fundamentally simple, then forwarding requires few decisions and the high-speed hardware is cheaper. If the core is complex, then forwarding requires more decisions and the high-speed hardware is more expensive. This is the reason that the core should be kept simple. The core should do IP routing and bulk QoS,and nothing else.
Apparently Foundry has built a router based on this principle. -----------------------------------------------
No arch_1, Foundry has built this router based on their MG8 enterprise switch. Thanks for playing.
Your concept of the dumb core of just switching (it is not "IP routing" except for control plane and maybe multicast traffic, most traffic is usually LSR function if you want dumb core) and QoS is incorrect. What about the very complicated LDP and RSVP protocols with multiple flavors of Fast Reroute, Diffserv aware TE, CSPF, BW reservation, LDP Filtering, LDP in RSVP tunneling, upcoming IPRR draft, LDP fast next-hop resolution drafts, IGP protocol extensions to support new features, Multicast PIM, MSDP, SSM, etc.? Not to mention that all of these features need to interoperate with other vendors, scale and be very stable.
That is not what i would call a dumb core and most of it is standard TODAY on any major Tier 1.
PS - a Router is a Router. They have to do many complicated things regardless of their role and have a large variety of interfaces (Look at all the features and interfaces on a 2600 router). The only way to make a real router cheap is to design and build it in India or China. The real cheap core router will come from Huawei, ZTE or someone like that... certainly not Foundry.
re: Foundry Strikes at the Core One comment to the PS here....
There is certainly a way for anybody to make a "cheap router" and that is to lower the price. It is not clear to me that the BOM cost of a Chinese Router will be less than that of an American Router. However, the R&D spend can be less (because of fewer $$ per engineer) thus price can be the lower to reach equivalent net margins.
One would think if you were not either Juniper or Cisco that this price drop would be the only way to enter the market.
Comment on OSMINE. They probably have to do TIRKS, which is no real work on the equipment and is very inexpensive and fast. Simply a way to simplify the warehousing of equipment.
re: Foundry Strikes at the CoreJust curious. I see mention that the T640 is a competitive product as well as the GSR 12k. What about the Cisco 7600? Certainly seems that while it appears the 7600 is a much more flexible chassis with the available interface types, I don't see it compared. Is there any reason it's not compared?
It supports WAN PHY which allows you to plug into SONET infrastructure and get the benefits of APS (I presume they support) which gives you the circuite quality service providers claim they need. This comes at the cost of additonal SONET framing encapsulation which means more bits on the 10G egress then the ingress (packet loss). Wonder how this box is at buffering...
re: Foundry Strikes at the CoreThank you, arch. It still sounds to me as though a fast, cheap, hardware-oriented router is simple-minded, no matter how you slice and dice the forwarding information. I keep reading about how L2 and L3 functionality is somehow merging. Perhaps this is the vehicle. Perhaps that is also why FDRY is looking at tier-2 service providers.
re: Foundry Strikes at the CoreI've been down this space before, so... 1. I agree with seven on TIRKS vs OSMINE. 2. Buffering is irrelevant for Enet vs SONET -- every IP router has to do this sort of "rate matching" thing, even if all interfaces are the same Network traffic just has to deal with it and back off. Conventional wisdom says 2 x RTT is the optimal amount of buffering. For the internet, this is between 200 - 500 mS. Summary: if it works in any scenario, it will still work in the core. 3. Pricing from American & Asian companies: A company which sells into multiple markets can drop their margins in one market, and sustain that for quite a long time. 4. Hardware routing -- LR should know better than to be caught in that bull. Everyone moved to HW forwarding >5 years ago. The current rage is network processors, which tend to be more expensive, but are more flexible and have reached a performance level which makes them feasable for 10G interfaces. Even with an NP, Foundry's argument falls flat. 5. Cisco 7600 vs others: Cisco 7600 hasn't been a core box. It just isn't big enough. (the PPS rate is too low) and it is marketed as an edge box.
re: Foundry Strikes at the CoreOne would think if you were not either Juniper or Cisco that this price drop would be the only way to enter the market. ----------------------------------
That may not be the only way ,but sure it works. Didn't Huawei just prove that ?
For other ways , more innovative business models are required .Any takers ?
arch_1, good primer. What I read is: data plane has limited autonomy (like the states) while control plane has overall control and will interfere when necessary (like the fed). Does that mean one needs a cpu plus an npu/asic to build the router, or only one chip can do the job? If so which one? TIA.