Hello Olivier,

On 17/04/2013 12:28, Olivier Bonaventure wrote:
> Hello,
> 
> Here are some additional comments on the above draft. Section 4 suggests
> the utilisation of a hash to perform the load balancing based on the
> source-address/flow label pair. Hash functions are often used in load
> balancing applications. I think that it could be useful to point the
> advantages and the drawbacks of using hash functions for load balancing
> purposes.
> 
> The main advantage of hash functions is that they usually exhibit the
> avalanche effect, i.e. a small change in the input causes a large change
> in the output of the function. Simulations show that this effect is
> important to correctly balance real traffic.
> 
> However, with hash functions, this advantage comes with a drawback. It
> is difficult to predict the set of inputs that would provide a specific
> output. 

That is an advantage from the security point of view, since it
prevents an attacker from biasing the load balancing in order to
focus a DOS attack on a single server. I believe that's discussed
in the flow label spec.

> Being able to predict which decision will be taken by a load
> balancer is important for monitoring applications for example. If a
> network operator wants to verify the round-trip-time between two hosts
> when there is a load balancer in between, he/she should take into
> account this load balancing when defining his/her probes. A similar
> situation happens when a network operator wants to use traceroute to
> detect block holes on a load-balanced path.

True, but there's likely to be little sympathy from the load balancer's
operator, for the reason just given.

> It should be noted that there exist functions that exhibit the avalanche
> effect without being one-way functions like the classical hash
> functions. If used in load-balancers, these functions would provide both
>  good load balancing and predictibility which is desired for many
> monitoring applications. A recent paper shows how such load-balancing
> can be performed efficiently by using block ciphers instead of hash
> functions. The solution proposed in this paper could be adapted to
> utilise the IPv6 flow label :
> 
> http://inl.info.ucl.ac.be/publications/revisiting-flow-based-load-balancing-stateless-path-selection-data-center-networks

We'll have a look, thanks for the reference.

I believe, from the research I had to do personally when working on
this draft, that there is scope for an RFC giving a general overview
of load balancing. That would certainly have been a precious reference
for the current draft.

Regards
   Brian

> 
> Best regards,
> 
> 
> Olivier Bonaventure
> 
_______________________________________________
Int-area mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/int-area

Reply via email to