|
Hi,
On Thu, 13 Mar 2003 jpcl@xxxxxxxxxxxxxx wrote:
> I want to try KTCPVS as a load balancer with sticky session (cookie) support.
> Well, I was theoretically thinking on this and I don't see how this
> problem is solved:
>
> I use LVS-DR. So packets go to Director, are forwarded to Realserver, and
> then go direct to outside router.
> If cookie management is at director level, and packets go in one-way only,
> then:
>
> 1 - Client send request {req}. Request has no cookie attached.
> 2 - Director sees no cookie, forwards to RS1
> 3 - RS1 gives cookie COOKIE1
> 4 - Packet goes out with {COOKIE1, reply}
>
> 5 - Client send new request {COOKIE1, newReq}
> 6 - Director sees COOKIE1, but has no mapping information of it ?!?
>
You can use the persistence feature of LVS-DR to solve this cookie
problem. If the virtual service is specified as persistent, all the
requests from a single client (per IP address) will be redirected to the
same server selected for the first request.
> This doesn't seems to be a problem specific to KTCPVS, but to all
> arquitectures who do not use a "two-way" balancer.
> Maybe the solution is completely different, and the balancer itself
> inserts the control cookie?
>
> You'ld have
> 1 -
> 2 - Director sees no cookie, forwards to RS1 as {DIR_COOKIE1, req}
> 3
> 4 - Packet goes out with {DIR_COOKIE1, COOKIE1, reply}
> ...
>
> Is this the racionale behind KTCPVS phttp behaviour?
>
Yes, KTCPVS can remember the cookies and send requests according to the
cookies, but current KTCPVS code doesn't support. We are thinking about
adding this support.
However, with content-based scheduling module, you may use the rule to
assign some stateful requests to the same servers and stateless requests
to other servers.
> Thank you
> Joao Clemente
>
Regards,
Wensong
|
