Scalability issues in the Internet routing system
vijay gill
vgill at vijaygill.com
Tue Oct 18 19:14:03 UTC 2005
Andre Oppermann wrote:
> vijay gill wrote:
>> Moore's law for CPUs is kaput. Really, Moore's Law is more of an
>> observation, than a law. We need to stop fixating on Moore's law for
>> the love of god. It doesn't exist in a vacuum, Components don't get
>> on the curve for free. Each generation requires enormously more
>> capital to engineer the improved Si process, innovation, process,
>> which only get paid for by increasing demand. If the demand slows
>> down then the investment won't be recovered and the cycle will stop,
>> possibly before the physics limits, depending on the amount of demand,
>> amount of investment required for the next turn etc.
>
> Predicting the future was a tricky business ten years ago and still is
> today. What makes you think the wheel stops turning today? Customer
> access speed will no increase? No more improvements in DSL, Cable and
> Wireless technologies? Come on, you're kidding. Right?
Missing the point. We can deal with increased speeds by going wider, the
network topology data/control plane isn't going wider, THAT is where the
moore's observation was targeted at.
>
>> Also, no network I know is on the upgrade path at a velocity that they
>> are swapping out components in a 18 month window. Ideally, for an
>> economically viable network, you want to be on an upgrade cycle that
>> lags Moore's observation. Getting routers off your books is not an 18
>> month cycle, it is closer to 48 months or even in some cases 60 months.
>
> When you are buying a router today do you specify it to cope with 200k
> routes or more? Planning ahead is essential.
>
And we're paying for it. But again, assuming that the prefix/memory
bandwidth churn can be accommodated by the next generation of cpus. I am
not going to throw out my router in 18 months. Its still on the books.
>> Then we have the issue of an memory bandwidth to keep the ever
>> changing prefixes updated and synced.
>
> Compared to link speed this is nothing. And nowhere near to memory
> bandwidth.
Each update to and fro from memory takes cycles, and as the routing
tables become bigger, the frequency of access to the memory for keeping
the system in sync impose a larger burden. This is orthogonal to link speed.
/vijay
More information about the NANOG
mailing list