[c-nsp] etherchannel load-balancing & WS-X6708 issue

Jiri Prochazka jiri.prochazka at superhosting.cz
Fri Sep 9 07:33:14 EDT 2011


Hi,

I'm trying to establish 40 Gbps redundant cirle using a pair of 6500 and
7600 boxes equipped with X6708 cards as shown below -> 
 _________              _________
 |        | Te1/2 ------|        |///
 |  core  | Te1/1 ------| edge-2 |//
 | 7600-1 | Te1/5 -Po1--|        |/
 |_____ __| Te1/6 ------|________|
   3 4 7 8
   | | | |
   | Po3 |
   | | | |
   | | | |
 __2_1_5_6__             _________
 |         | Te2/3 ------|        |
 |  core   | Te2/4 ------| edge-1 |\
 | 6500-1  | Te2/7 -Po1--|        |\\
 |_________| Te2/8 ------|________|\\\

There is one WS-X6708 in each core box which should be dedicated to this
'circle'. (Besides it there are of course other 670x for custs). I believe I
have quite good knowledge of 6708 DFC architecture and it's limitations (16
Gbps for pairs of ports, 20 Gbps for each two pairs).

Primary box is 6500-1 (several hunderd Vlans and SVI's) and under standard
circumstances Po1 is the only path utilized, but even if I freak out, I am
not able to push more than aproximately 25 Gbps from 6500 over Po3 to
edge-2. My initial guess was bad port assignment and therefore not utilized
local switching..

The pairs which can utilize the local-switching (so 16*4..64 Gbps) should be
following

2-3, 1-4, 5-7, 6-8

The major problem which I can not beat (and maybe it's a dead end for me) is
the system how IOS decides which ports will be assigned to 0-7 ID's used in
etherchannel load balancing algorithm. As I observed, it absolutely depends
on a sequence of adding/removing ports from/to an etherchannel.

Let's say I have created 40Gbps Po3 on 6500 and IOS made following divison
->

Te2/2 would be used for traffic with RBH 0x0, 0x5
Te2/1 would be used for traffic with RBH 0x3, 0x7
Te2/5 would be used for traffic with RBH 0x6, 0x2
Te2/5 would be used for traffic with RBH 0x4, 0x1

! BUT !

Po1 on 7600-1 has absolutely different traffic pattern computed!

Te1/2 would be used for traffic with RBH 0x7, 0x2
Te1/1 would be used for traffic with RBH 0x6, 0x1
Te1/5 would be used for traffic with RBH 0x3, 0x4
Te1/5 would be used for traffic with RBH 0x0, 0x5


And everytime I shut/unshut some bundled port, this changes :-). As soon as
I don't have the same values on both boxes, local switching is not used..


I hope I expressed the least bit comprehensibly..



Any thoughts are really appreciated!




Regards,


Jiri Prochazka



More information about the cisco-nsp mailing list