At 1:08 AM +0000 6/9/03, The Road Goes Ever On wrote:
>
>
>
>or to put it another way, why bother when memory and CPU is relatively cheap
>( you DO use 3rd party memory, don't you ;-> )
>
>As I tell all my customers, it doesn't hurt to max out the memory. Never can
>tell when you will need it. ( and it helps me retire quota )
>
>As I say when I want to yank Priscilla's chain, design is dead. This kind of
>work is irrelevant.
Ah, but what is design?
Seriously, in all but the very largest enterprises, your advice is
quite reasonable. It wasn't always the case that adding 500 or 1000
routes was insignificant, especially when the 7000's silicon
switching cache held only 500 or 1000 routes.
Right there, however, is an issue I'd call "operational design" or
"implementation engineering." The issue, even then, wasn't RAM
usage, but cache thrashing.
And so it is, even today, with small enterprise routers, or with
large carrier routers. Raw memory size is rarely the problem, but
route processing can be. For OSPF, an approximation of the load to
do a Dijkstra computation is:
(routes * routes) * log(routers)
of course, the Dijkstra is only part of the route computation,
dealing with intra-area routes only. The total computation then
grows linearily with the number of inter-area and external routes
that the area must consider.
Even more basic is how often this computation has to be repeated,
which comes back to the question of route stability. The Dijkstra
algorithm itself is 40-some years old, and faster variants are
available. There's no way we can get into subsecond convergence with
an unmodified Dijkstra and alternatives to hellos.
In the absence of additional processing that uses the same processor
[1], the limiting factors tend to be less the absolute number of
routes than the rate of route change and the level of interconnection
among routers (hierarchy decreases that). A major issue in the global
Internet is less the load on specific routers, than it is the overall
slowing of convergence and a greater tendency to blackholing.
Now, when you get not into "network design" but "protocol design,"
some of these factors that you could ignore re-raise their heads and
bite you. In a large ISP router, memory speed is a major limitation.
As a designer, you may indeed be limited by the cost of SRAM vs.
DRAM, or whatever is the fastest available memory. You will also be
limited by the forwarding fabric -- shared bus, as in the 7500, tops
out at around 2 Gbps with shared memory being somewhat but not
overwhelmingly more capable. Crossbar fabrics then become the basic
usable technology, but they have their own problems of a limited
number of points of connection.
[1] Knowing what other processing goes through the route processor is
important and also release- and platform-dependent.
Message Posted at:
http://www.groupstudy.com/form/read.php?f=7&i=70402&t=70299
--------------------------------------------------
FAQ, list archives, and subscription info: http://www.groupstudy.com/list/cisco.html
Report misconduct and Nondisclosure violations to [EMAIL PROTECTED]
