[alcatel-nsp] unbalanced traffic on GigEs in a LAG

[Ryan Landry]

On 2010-03-08, at 5:40 PM, Ryan Landry wrote:

> 
> On 2010-03-08, at 5:17 PM, Scott Weeks wrote:
> 
>> 
>> --- philxor at gmail.com wrote:
>> From: Phil Bedard <philxor at gmail.com>
>> 
>> I guess it's hard to know exactly which services are flowing across the 
>> link but if it's only acting as an LSR then it should be hashing based 
>> on the MPLS stack (Up to 5 labels), ingress port, and system ID.   Are 
>> there any potential really high BW services across the LAG?  With the 
>> IOM3/IMM you can have it hash on both the labels and the IPv4 header, 
>> but not on the IOM2.  
>> 
>> If there are services originating on the box there are all kinds of 
>> rules on how things are hashed on egress, but LSR it's pretty simple.  
>> --------------------------------------
>> 
>> 
>> We carry our internet traffic in a VPRN.  This is 99+% of all traffic.  
>> I'm guess I'm going to have to go with just 2 cases of bad load 
>> balancing.  The network goes something like this:
>> 
>> 
>> br1     br2
>> |       |
>> |       |
>> 7750----7750
>> |\       /|
>> | \     / |
>> |  \   /  |
>> |   \ /   |
>> |    X    |
>> |   / \   |
>> |  /   \  |
>> | /     \ |
>> |/       \|
>> 7750----7750====7750
>>            ^
>>            |
>> 
>> 
>> No links are full.  The 'middle' is 10G and the 'edges' are multi-GigE.  br1 has about twice the bandwidth as br2, so perhaps the label from br2 is hashed over one of the 2 GigEs (above the arrow) and the label of br1 is over the other  and it's just a case of bad load balancing.  More that 10 to 1, though, seems excessive.
>> 
>> 
>> Thank you everyone on and off-list for your help.  :-)
>> scott
> 
> well, you carry your inet in a vprn, so this may not be applicable, but you can included L4 in the hash algorithm.  i haven't tested it nor had need for it.
> 
> configure system l4-load-balancing
> configure port <id> eth load-bal include-l4|exclude-l4
> 
> other than that, i've really had no problems with load being split unevenly, be it ECMP or LAG, in the lab or otherwise.
> 
> does your dst inet traffic basically have one egress LSR?  ie: maybe you just don't have a lot of src/dst to hash against.
> 
> g'luck.
> 
> .rL

yeah, so you answered my question before i even answered it.  my apologies for not reading the whole thing thru before replying.

.rL