|
The only reason i found is that router/lb has also the IP of the lb and
for him the response can't be send by someone else than him, a tcp stack
/connection tracking problem in some sort.
If you use the director as default gateway for your RS in LVS-DR mode,
you need to patch your kernel with the shared forward patch, to be found
here:
http://www.ssi.bg/~ja/forward_shared-2.6.17-2.diff
Please report back if that works for you or not.
Effectively that work fine on a 2.6.18.
Thanks.
Should this be merged. I'm happy to try and push it in if it should
be in the main tree.
Among other things not to be shared here in public, I had the exact same
thought in the shower this morning :). It renders life so much easier
for people, because the whole triangulation or asymmetric forwarding is
just a tad bit too unusual for most people to care enough. And setting
up a load balancer for a project is most of the time just a nitty-gritty
and highly technical part of the whole process. Who wants to spend 80%
of the budget for 20% of the project? At least that's what I've been
confronted with in projects where IPVS was in discussion, as one of the
technologies to be used.
Regarding the political point of view: I'm not sure if Julian wants to
step up again against the whole netdev-crew for yet another "special"
feature that no one else in the networking world needs. Although, it
would be Horms' call after all.
Regarding the technical point of view: That patch is very non-intrusive
and only adds one branch (could even be marked unlikely()) to the FIB
frontend and this is certainly acceptable.
Julian, do you have any technical reasons that would warrant a veto to
the inclusion of your forward shared patch into the main linux kernel?
The time would be good now to push it to DaveM for 2.6.20, I believe.
<sidenote>
On a more promotional side, I'd like to mention that we should focus on
getting IPVS into a shape of a very well documented piece of framework,
that is easy integrable into an existing project. Forward shared is one
piece of the puzzle in my humble opinion, using VRRP is another. Most
commercial load balancers these days are not set up anymore using the
triangulation mode (at least in the projects I've been involved), it's
becoming more and more a router using well-understood key technologies
like VRRP and content processing. And after almost 10 years of my
involvement with load balancers I have to admit that no customer _ever_
truly asked or cared about the scheduling algorithm :). This is academia
for the rest of the world.
</sidenote>
Best regards,
Roberto Nibali, ratz
--
echo
'[q]sa[ln0=aln256%Pln256/snlbx]sb3135071790101768542287578439snlbxq' | dc
|
