From: Craig Partridge [mailto:craig@aland.bbn.com] Sent: Tuesday, April 10, 2001 4:36 PM
In message <Pine.BSF.4.21.0104101753540.98098-100000@overlord.e-gerbil.net>, "R ichard A. Steenbergen" writes:
Don't be absurd, I can walk into fry's and pick up a motherboard with 64bit/66mhz PCI, some Netgear GA620's, and all the other components for a 1GHz computer for under $1000.
OK, so your bus has 4.2 Gb/s of bandwidth. But, alas, you're in a PC so you have to copy each packet from the line card, into main memory, examine it, and push it back out to a line card. So each packet consumes twice its size in bus bandwidth. So 2 1 Gb/s line cards will consume 4 Gb/s backplane. Assuming you can run the PCI at full rate (which in my experience is a big big if), you can connect two Ethernets.
There are a few server-class motherboards that have dual PCI-66 buses. One on each bus and there's no problem. Incidently, those same MBs usually have dual or quad CPUs.
Incidentally, this isn't the full story either. You have to do a route lookup on each of those packets. That's typically 5 to 10 memory accesses... 5 memory access times 1 Mpps per gigabit times 2 gigabits is 10 million lookups per second or 100 ns per lookup. Allowing for time spent getting through the chip to the pins, you probably need 60 or 70ns DRAM, which is doable. Except, oops!, that completely consumes your memory bandwidth... where are you going to find the cycles to get the packets in and out?
Huh?! We haven't had the 60ns limitation for over three years. What time-vault have you been living in? PC100 RAM is error correcting at 10ns and PC133 runs at ~8ns. In fact, I challenge you to show me where you can even buy 60ns RAM these days. That's right up there with 3 GB HDDs, not available because they can't find customers for any thing that small/slow, except propriatary applications (they make you pay ... large). I just paid $104 for 256MB of ECC PC100 RAM, at the local computer store. The discount wasn't even worth the gas/hr to drive down to San Jose for (1989 Chevy Suburban, in San Jose rush hour).
PS: Side note, this illustrates where router vendors earn their bucks.
Router vendors earn their bucks by building routing/switching appliances. The fact that they manage to get corps to cough up $200K for $15K worth of hardware (COGm, generous) is yet another issue. I expect to see some serious price drops in the next 6 months. I also expect to see commodity hardware start making some serious in-roads (cause of the price-drop). COGr = 1/2 COGm, manufacturers cost of raw material (chips, labor, copper, and such) COGm = Cost of Goods manufactured (fully burdened) 2*COGm = The resellers cost (including mfg's margins) 4*COGm = The distributors price 8*COGm = Retail price (What you pay, full retail) [yeah, that's right. This means that, said RAM, cost $13 to make. econ 101]
Find a way to move data over each bus only once (double your bandwidth!).
Dual PCI66 bus.
Design your memory subsystems to keep packets and routing data separate (increase your memory bandwidth!).
PC133 ECC RAM (8ns).
Find a processor that doesn't waste cycles doing virtual memory (improve your memory access times!).
Dual/Quad 1Ghz PIIIs, anyone?
Oh yes, and then add hot board swap,
How many routers actually do this? Even when they do, it is more prudent to find the time to shut it down for 5 minutes. It's not like solid-state devices die all the time...no moving parts. Keep them cool and they'll go obsolete before they fail. Unless it's a combat situation (where you don't have time to swap anyway) there isn't much that warrants this level of maintenance. You're much better off with a redundant hot fail-over configuration so that you can bring one down for PM and replacement/upgrade.
a working BGP implementation (quick, where's Tony Li working these days:-)),
I actually like GateD. It ain't fancy, but it works.
a CLI,
Linux is handy. RT Linux might be better (maybe).
and a power subsystem for a CO,
If you must include that, then my choice is a Liebert GT3600.
and you're in business.
With current commodity price drops, in less than 6 months, the cost of doing this will fall well below $1000, retail. Considering potential market volume. A 100% proprietary hardware solution may not be feasible, at that price-point. They will *have* to use commodity hardware, just to keep up the margins.