[j-nsp] 802.3ad trouble (part II)

[Jorn Ekkelenkamp]

Chuck Anderson wrote:
> On Thu, Nov 09, 2006 at 04:45:35PM +0100, Jorn Ekkelenkamp wrote:
>> transfers (multiple downloads at the same time, from different source 
>> addresses) interface fe-0/1/3 is not used AT ALL (see below). This 
>> results in a maximum speed of 100mbit for the whole ae1 interface.
> 
>> # show configuration policy-options policy-statement load-balancing-policy
>> then {
>>      load-balance per-packet;
>> }
> 
> "load-balance per-packet" is a misnomer, held over from the original 
> Internet Processor ASIC.  On Internet Processor II ASICs it really 
> means "load-balance per-flow".  It appears that the different source 
> addresses you are choosing to test with are being hashed to the same 
> physical interface.  You can add layer 4 information to the hash:
> 
> set forwarding-options hash-key family inet layer-4

I've switched on a few extra options, just to be sure (thanks to Niels 
from AMS-IX):

# show configuration forwarding-options
hash-key {
     family inet {
         layer-3 {
             destination-address;
             protocol;
             source-address;
         }
         layer-4 {
             destination-port;
             source-port;
             type-of-service;
         }
     }
}

This does seem to have some effect, both links show output packets now:

     Link:
       fe-0/1/3.500
         Input :             0          0             0            0
         Output:       2542480          0    3812023009            0
       fe-1/1/3.500
         Input :       1350699          0      70763948            0
         Output:        136271          0     198056537            0

The switch still does not send packets back to the Juniper on one link 
though.. I guess we'd need to turn on similar features on the switch in 
order for this to work.  Too bad the switch does not seem to support 
tweaking layer[34] stuff like that :-(

The only thing I don't fully understand yet, is that I'm still not able 
to push more than ~92mbit over the link (from the Juniper --> ae1 --> 
server), since the Juniper should now be more or less loadbalancing 
everything it sends to the ae1.

Thanks!