routeing table size gotcha (IPv6 ?)
Peter Galbavy
otherwise. Having innocently tried to upgrade our router kernel to
2.6-current (to support the if_dc driver for performance) I suddenly found
that the kernel was running out of memory.
Where in the past, carrying 72,000 odd routes (145,000 entries) used about
11M of routetbl entries (150 byes or so each), with the new kernel I got to
60,000 routes (120,000 routetbl entries) for 16M of space (about 260 bytes
an entry).
With the growth of addresses from 4 to 16 bytes, is this major size increase
due to (a) alignment issues, (b) prefix/netmask/etc. additions or (c) me not
knowing what is going on ?
If there is someway to optimise the FIB for full-Internet route carrying
routers, I think that this would be well received.
Regards,
--
Peter Galbavy
Knowledge Matters Ltd.
http://www.knowledge.com/
ito...@iijlab.net
I assume you are carrying bunch of IPv4 routes, not IPv6 routes.
I don't think IPv6 integration affected your problem. IPv4 routing
entries are kept unchanged. sys/net/route.c handles address families
separately, and it does not require change in size even if we integrate
IPv6 into the kernel.
Just in case, what happens if you remove "options INET6" from your
kernel configuration? It would decrease kernel code size and
may make things better for you.
itojun
Peter Galbavy
> otherwise. Having innocently tried to upgrade our router kernel to
> 2.6-current (to support the if_dc driver for performance) I suddenly found
> that the kernel was running out of memory.
Well, it's not IPv6, so that's OK. struct rtentry seems to have grown along
the way from 2.5. As per a private mail, is there any point considering
trying to optimise (for memory) the way that route entries are handles for
large-route based installations ?
I haven't looked at Steven's Volume 2 in a while, but it may make my reading
list in the very near future.