below...
On 2011-06-07 11:22, Bob Hinden wrote:
> Ran,
>
> On Jun 6, 2011, at 9:39 AM, RJ Atkinson wrote:
>
>> I would be reluctant to advocate use of a hash algorithm
>> as computationally expensive as MD5() for generating an
>> IPv6 Flow Label.
>>
>> My sense is that a thoughtful implementer likely would
>> choose the least computationally expensive algorithm
>> that would suffice to generate a Flow Label value
>> suitable for load-balancing purposes.
>>
>> As a use case, one might consider some sort of content server
>> with very large numbers of short-lived TCP sessions. If an
>> expensive algorithm (e.g. MD5()) were used in that case,
>> then the effort to compute the flow label might be am
>> interesting fraction of "new TCP session creation" overhead.
>> In turn, this might impair the scalability of that content
>> server.
>
> I suppose one could precompute a bunch ahead of time, but I agree with your
> comment. It's a good point.
>
>> In any event, I'm quite comfortable with Brian's current
>> approach to document a possible algorithm that ought to
>> suffice without mandating any particular algorithm.
>> This really is a situation where the details are best
>> left to implementers (since no interoperability issue
>> would arise from different implementations using different
>> algorithms).
>
> I also agree.
>
I sort of agree, if we're talking about my proposed replacement algorithm
However, I would want to weaken the language in the appendix to make it clear
that better algorithms can and will be found - I've had to break off my
experiments for a few weeks due to travel, but there is a lot of scope
for improvement. Anyway, let's see what the AD thinks.
Brian
--------------------------------------------------------------------
IETF IPv6 working group mailing list
[email protected]
Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
--------------------------------------------------------------------
