[c-nsp] memory-size iomem (7200 series)

Robert E.Seastrom rs at seastrom.com
Thu Oct 13 08:16:18 EDT 2005


Andrew Fort <afort at choqolat.org> writes:

> Robert E.Seastrom wrote:
>
>> Yes; I'm not joking; I think if you ask around you'll find that there
>> are a *lot* people who are still running the pre-ST-merge code because
>> of the little bit of extra memory it saves.  Handy when you have a
>> bunch of NPE-300s, RSP-4s, or original GRPs lying about.  You can
>> still take full BGP views in those you know...
>> 
>> Yes, it's lacking a lot of features found in newer code.  When you're
>> not using any of them though, it ends up not buying you anything.
>
> one could rightly argue you're using "the new features" if you're using 
> BGP at all... the changes have really made a difference in the last 12 
> months with all that growth in table size we're seeing again, and 
> arguably have extended the life of some of the slower processor units if 
> they have to deal with large views.  memory aside, of course :)

s/large views/many partial views or big ibgp mesh/.

In degenerate-case small multihomed networks (and I run a few of those
under contract, nice work if you can get it), the CPU is still quite
happy, and it's all about the memory.  It's a bitter pill for people
to swallow that they have to cough up bucks for an NPE-400 when
they're running 15 mbps total in+out on a heavy day.  Obviously,
they'll have to do it eventually given routing table size growth...
but logging into one particular VXR that's running pre-st-merge code
with two full views and an ibgp view from its partner (CEF on,
soft-reconfig off), we see 58 megs free.  Another router, similar
architecture, same number of views (even same distribution, ibgp
vs. ebgp), running 12.3T (both "p" loads) is running with a scant 5
megs free.  Now, I realize that's not the best test, that we should be
looking at 12.0S post merge to see what the real cost is, because
12.3T has a lot more "stuff" in it and so less memory for routes
regardless of how they are stored...

                                        ---rob




More information about the cisco-nsp mailing list