Juniper MX10 and dual stack BGP
saku at ytti.fi
Thu Jan 31 07:25:58 UTC 2013
On (2013-01-30 21:06 -0500), David Miller wrote:
> According to Juniper, the MX uses separate memory for v4 and v6.
Where do they state this? MX is ambiguous, what matters is linecard HW.
> The numbers that I have seen for MX80 are:
I.e. trio. No. Trio uses flat RLDRAM, and any IPv6 route installed will mean
less IPv4 routes can be installed.
# show luchip 0
RLDRAM: 576 Mb by 4 devices at 533 MHz.
DDR3: 1024 Mb by 2 devices at 733MHz/CL10.
I.e. 576 * 4 * 8/9 * /8 = 256MB
# show jnh 0 pool summary
Name Size Allocated % Utilization
EDMEM 33554432 15238492 45%
IDMEM 323584 159278 49%
OMEM 33554432 33079304 98%
Shared LMEM 512 67 13%
RLDRAM == EDMEM + IDMEM
DDR3 == OMEM (mobilenext, flow stuff)
TCAM == not used
(E series MPC have twice the DDR3)
> v4 FIB 1mil
> v4 RIB 4mil
> v6 FIB 512k
> v6 RIB 3mil
I believe these are the numbers JNPR test with and are mutually exclusive.
256MB is fuckton of memory for lookup, if they can use it fully, but due to
caching reasons, some stuff might be mirrored to each bank to accelerate
lookup, so I'm unsure how far you can scale Trio.
More information about the NANOG