|
Julian Anastasov wrote:
>
> On Fri, 28 Jul 2000, Chris Anderson wrote:
>
> > So, I guess the question is, how do I go about sending stuff to the
> > realservers by splitiing it up at the packet level, rather than the
> > connection level?
>
> Good question! This is possible to implement. But be more
> specific:
>
> - protocol (UDP)? LVS talks only TCP and UDP for now.
TCP/IP
> - why the restriction for only one source?
That's covered by the NDAs :-/ I wish I could say more, as it would
help with my explanations.
Think of it as, oh, a proxy that delivers all of its connections to us.
> - what compression?
Stripped headers that we put back together at our end stuff like that.
> - RFC?
I'm not sure of any RFC that covers what we're doing; it's all bleeding
edge from where I stand.
> - more details and pictures
>
> I see this as a service (and so an entry) flag. Of course,
> if there is a good reason for such flag without raising other
> problems. Scheduling on each UDP packet without creating entries?
> Or one for each source? Any application that requires such feature?
The incoming packet stream needs to be balanced among several real
servers, which do their own processing on the data, and send it on
elsewhere without any further intervention on the part of the director.
Any return traffic is not related to the incoming connection, so storing
info about the connection is not required (as best as I can figure out
at this point). A scheduling system such as weighted least realserver
load would be nice, too. (I'm seeing a lot of custom script writing here
:-/ )
Just as an aside: is it possible to change the weighting on the fly when
using weighted whatever scheduling, without having to rewrite the config
file and restart the director? Maybe based on realserver load or some
other metric?
Cheers, Chris.
--
"This calls for a very special blend of psychology
and extreme violence." -- Vyvyan, "The Young Ones"
|
