[c-nsp] multipath BGP not balancing equally.

[David Hughes]

Hi

But seeing as the OP indicated that one of the circuits was 2GB  
*underutilised* you'd be looking for 3 src/dst pairs that were all  
doing 2GB to get this situation.  It's looking pretty unlikely that  
this is a hashing issue.


David
...

On 06/08/2009, at 6:23 AM, Rodney Dunn wrote:

> Ah...good one. If the sources were not random enough and it's NAT'ed  
> to one external ip you could really be multiplexing flows with NAT. ;)
>
>
>
> Dean Smith wrote:
>> Would agree that volume is rare between 2xIP addresses but we have  
>> something similair although on not quite the scale.
>> We NAT a very large organisation to the Internet. They have a large  
>> number of disparate sites that all do their own AV updates. All the  
>> PCs download at the same time in the evening and we generate about . 
>> 75 Gb/s of traffic between our external PAT address and the AV  
>> download site for a good couple of hours. If we had a bigger  
>> internet pipe it would be a higher figure. (for less time of course).
>> Dean
>> ----- Original Message ----- From: "Rodney Dunn" <rodunn at cisco.com>
>> To: "Mikael Abrahamsson" <swmike at swm.pp.se>
>> Cc: "Cisco" <cisco-nsp at puck.nether.net>
>> Sent: Wednesday, August 05, 2009 2:19 PM
>> Subject: Re: [c-nsp] multipath BGP not balancing equally.
>>> For small flow combinations you are right. btw, it would be just  
>>> L3 src/dst flows by default unless the L4 port option is enabled.
>>>
>>> I thought about there being a single flow causing the difference  
>>> that would be hashing down one of the paths. But 2G, while not  
>>> impossible, typically isn't used between two ip addresses. It's  
>>> something to check though for sure.
>>>
>>> Rodney
>>>
>>>
>>>
>>> Mikael Abrahamsson wrote:
>>>> On Tue, 4 Aug 2009, Rodney Dunn wrote:
>>>>
>>>>> That's usually caused by routes not being the same on the paths.
>>>>
>>>> It was my understanding that this usually was caused by not  
>>>> having enough L4 flows to loadshare on...? Ie if you have 100 TCP  
>>>> flows and 4 paths, then it's not enough flows to get good load  
>>>> share on, but if you instead have 10k flows and all of them are  
>>>> low-speed, then the odds of them being equally load shared is  
>>>> much better?
>>>>
>>> _______________________________________________
>>> cisco-nsp mailing list  cisco-nsp at puck.nether.net
>>> https://puck.nether.net/mailman/listinfo/cisco-nsp
>>> archive at http://puck.nether.net/pipermail/cisco-nsp/
>>>
>>> __________ NOD32 4306 (20090804) Information __________
>>>
>>> This message was checked by NOD32 antivirus system.
>>> http://www.eset.com
>>>
>>>
> _______________________________________________
> cisco-nsp mailing list  cisco-nsp at puck.nether.net
> https://puck.nether.net/mailman/listinfo/cisco-nsp
> archive at http://puck.nether.net/pipermail/cisco-nsp/