RE: verio arrogance

Mailman List Mirror (Read Only)
1

Getting back to the more original thread.

Is there any need to keep the routing table to a smaller size. Since in theory, it creates suboptimal routing. And considering the new routers out there today should be able to handle it. Considering verio is using junipers, and they pride themselves on handling a tremendously large table. Why should we shoot for a 100,000 route table instead of 500,000 if it does not impact performance?

I do understand that the 100,000 might be that actual 'installed best routes' and that the routers might in fact be dealing with a much larger route table. That might be an issue. But certainly 100,000- 500,000 installed routes, is that a problem for large backbones with high end routers?

My only consideration might be the small multihomed ISPs with 2-3 providers with full BGP feeds and cisco 4000s (256meg ram). I saw one last week. I might be concerned at that level.

I'd love to hear feedback. It would then justify filtering...or not.

David

2

Is there any need to keep the routing table to a smaller size. Since
in theory, it creates suboptimal routing. And considering the new
routers out there today should be able to handle it. Considering
verio is using junipers, and they pride themselves on handling a
tremendously large table. Why should we shoot for a 100,000 route
table instead of 500,000 if it does not impact performance?

  When you are talking about BGP reconvergance when a router
crashes (oh wait, they would never crash :wink: or is upgraded it takes
a lot longer to advertize 500k routes than 100k routes. Even
with a really-fast processor it obviously takes more time to do
route lookup in doing best-path computations with 100+ ibgp
peers.

  Then you start to talk about the memory footprint of 500k
prefixes, once you start to include received-side communities
as well as your new communities you've tagged on. With
route-refresh it's not that bad, but with soft-reconfiguration enabled
it may cause a bit more memory to be used.

I do understand that the 100,000 might be that actual 'installed best
routes' and that the routers might in fact be dealing with a much
larger route table. That might be an issue. But certainly 100,000-
500,000 installed routes, is that a problem for large backbones with
high end routers?

  If you venture a guess and say that most "large" networks
originate about 5% of the 100k prefixes must be advertized (see
peering discussion about minimum routes to advertize awhile back)
that numer of prefixes is increased to 25k prefixes. Then if you
prefix-filter your customers, you're talking about 5X increased
nvram/config requirements.

My only consideration might be the small multihomed ISPs with 2-3
providers with full BGP feeds and cisco 4000s (256meg ram). I saw
one last week. I might be concerned at that level.

  "back in the day when full routes would fit in 64m ram".
obviously the smaller providers have a bit more of a challenge as
they tend to not have support contracts, and it can be a bit
tougher to justify router memory.

I'd love to hear feedback. It would then justify filtering...or not.

  Think about the "7007" and other cases whereby someone
announces a large set of routes they should not be.

  There have been numerous cases of this in the past and as
a long as it's possible to easily leak routes incorrectly due to
not filtering customers closely, etc.. it will continue to happen.

  - jared

3

table. Why should we shoot for a 100,000 route table instead of 500,000
if it does not impact performance?

Convergence time?

- kurtis -

4

I think we are at the point where the vast majority of backbone routers can
handle 200K+ routes, at least in terms of memory. The interesting point we
are getting to, is that the most popular router in the world for multihoming
can't handle the routing table. I'm referring to the Cisco 3640, which has
largely supplanted the venerable 2501 as the low-end multihomer's edge
router of choice.

With a reasonable number of features turned on (i.e. SSH, netflow, CEF), the
3640 can't handle two full views anymore, due to it's limitation of 128MB.
While this may be a good thing for Cisco's sales numbers, in this winter of
financial discontent, I wonder how this is effecting the average customer,
and what is generally being installed to replace the 3640s.

- Daniel Golding

5

I think we are at the point where the vast majority of backbone routers can
handle 200K+ routes, at least in terms of memory. The interesting point we

I've not tried it but I doubt the ability of any Cisco router up to and
including 7500 to handle that many routes....

Btw, sketchy partial BGP tables seems to be the most common workaround to memory
limitations.

6

If a 3640 customer can't handle multiple full views, why can't they
filter some junk /24s themselves? This isn't really a good enough reason
for backbone providers to do the filtering.

As for the convergence time argument, the limiting factor is CPU time,
not the number of routes or amount of data exchanged (though obviously
more routes == more cpu). In the core, is there really that big a
difference between 93k and 113k? On the borders, how much cpu time is
saved vs how much cpu time is burned doing the filtering?

Which leaves us with the question of, are there still MSFC1's or other
devices with 128mb memory limits in these networks which are hurting at
113k? Is there actually a legitimate technical need to filter off 20k
routes, or are the people doing it stuck in a mental time warp from the
days when it was a necessity?

Or, is it really just people trying to do the "correct" thing? If you see
"almost" no change in connectivity after removing 20k of cruft, and the
very few people who are broken are the ones who needed attention called
to their poor route announcing skills anyways, maybe it's a good thing
for reasons other than router performance?

7

I think we are at the point where the vast majority of backbone routers can
handle 200K+ routes, at least in terms of memory. The interesting point we
are getting to, is that the most popular router in the world for multihoming
can't handle the routing table. I'm referring to the Cisco 3640, which has
largely supplanted the venerable 2501 as the low-end multihomer's edge
router of choice.

With a reasonable number of features turned on (i.e. SSH, netflow, CEF), the
3640 can't handle two full views anymore, due to it's limitation of 128MB.
While this may be a good thing for Cisco's sales numbers, in this winter of
financial discontent, I wonder how this is effecting the average customer,
and what is generally being installed to replace the 3640s.

If a 3640 customer can't handle multiple full views, why can't they
filter some junk /24s themselves? This isn't really a good enough reason
for backbone providers to do the filtering.

That was my thinking also. I would imagine a lot of customers what a full route view, it's what they are paying for especially if they are an ISP or multihomed large customer. They should have their own policies then.

As for the convergence time argument, the limiting factor is CPU time,
not the number of routes or amount of data exchanged (though obviously
more routes == more cpu). In the core, is there really that big a
difference between 93k and 113k? On the borders, how much cpu time is
saved vs how much cpu time is burned doing the filtering?

I would assume a flapping session with a large backbone would cause much higher load time and stress on the router then simply a large table. It's the reason why some backbones have Draconian route dampening policies, and rightly so. I would love to see some engineers from vendors weight in on this (did I just say that?). Most brag that they can handle large tables without a problem. A good question might be, if a large backbone started flapping 150,000 routes, what would that do to the peers. Perhaps a better issue much be CPU usage of complex route filters on large tables, as a limitation on performance.

Which leaves us with the question of, are there still MSFC1's or other
devices with 128mb memory limits in these networks which are hurting at
113k? Is there actually a legitimate technical need to filter off 20k
routes, or are the people doing it stuck in a mental time warp from the
days when it was a necessity?

Or, is it really just people trying to do the "correct" thing? If you see
"almost" no change in connectivity after removing 20k of cruft, and the
very few people who are broken are the ones who needed attention called
to their poor route announcing skills anyways, maybe it's a good thing
for reasons other than router performance?

Interesting thought is, there are probably a great many engineers on this list that have /24s at their home, that dont enjoy being filtered. Some of us just get tired of reIPing our servers.

dave

8

I couold be wrong, but I thought the 3640 had 256 mb of ram.

9

I believe the 3660 can do 256M. the 3640 can only
do 128M

http://www.cisco.com/univercd/cc/td/doc/pcat/3600.htm#spec