[Beowulf] ethernet bonding performance comparison "802.3ad" vs Adaptive Load Balancing

[Eric Thibodeau]

Martin Siegert wrote:
> On Wed, Sep 17, 2008 at 07:51:01PM -0500, Rahul Nabar wrote:
>   
>> On Wed, Sep 17, 2008 at 7:23 PM, Eric Thibodeau <kyron at neuralbs.com> wrote:
>>     
>>> Rahul Nabar wrote:
>>>
>>> On Wed, Sep 17, 2008 at 4:05 PM, Eric Thibodeau <kyron at neuralbs.com> wrote:
>>> Well, I don't have "bondable" hardware so I'm really interested in how you
>>> technically manage this one at the end.
>>>
>>>       
>> The more I do this, the more I get this uneasy feeling that this
>> hasn't been done much before? :) Not too many guides that bond for
>> bandwidth-aggregation. None at all for strict peer-to-peer
>> bandwidth-aggregation. Am I trying to do the impossible?
>>
>> Most people seem to use bonding for fault tolerance or a one-to-many
>> communication pipe.I really need more anecdotes and comments from
>> other guys who successfully use bonding.
>>     
>
> It is my understanding that 802.3ad forbids what you want to do:
> running a single stream over more than one link; 802.3ad requires
> that all packets are guaranteed to be delivered in order.
>
> It is my impression that the standard was not written with HPC in
> mind: it addresses the scenario of running many streams over a few
> links, i.e., load balancing (and HA). 
>
> This does not mean that you cannot do what you want: you need to use
> round-robin mode (which AFAIK is still the default under Linux;
> easy to test with crossover cables).
>
> - round-robin mode violates ("is an extension to") the 802.3ad
>   standard because it does not guarantee in-order delivery.
>   In my experience this is irrelavant in a cluster environment:
>   often a single switch, no multiple hops, no routers - out of
>   order delivery is very rare and has very little impact on
>   performance when using round-robin mode (we have done silly tests
>   like one host with 4 GigE interfaces, one with 3 and still got close
>   to 3Gbit/s).
> - most switch vendors do not support round robin mode - the only one
>   that I know who does is Extreme (please correct me!).
>   You can get around that problem by using a separate switch for each
>   leg, but that requires that each host has the same number of interfaces
>   for that bonded network. E.g., you cannot have a host with a single
>   10GigE card and another host with 4 1GigE cards.
>
> Cheers,
> Martin
>
>   
Wooo...I think I'll explore this then with my new machines, link 
aggregation on the head node will come in handy for diskless nodes ;)

Eric

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20080918/f001fb0b/attachment.html