LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: DRWS-a layer 7 load balancer

To: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Subject: Re: DRWS-a layer 7 load balancer
From: 8923517 <gis89517@xxxxxxxxxxxxxxx>
Date: Wed, 24 Apr 2002 21:10:42 +0800 (CST)
On Wed, 24 Apr 2002, Wensong Zhang wrote:

>
>
> Hello,
>
> On Wed, 24 Apr 2002, 8923517 wrote:
>
> > > On Mon, 15 Apr 2002, motse wrote:
> > >
> > > >     Yes. At real server have another kernel module that can rebuild
> > the
> > > > connection base on the packet that contains URLs. The DRWS-filter is
> > the
> > > > module at real server. You can see the mechanism in the slides.
> > > >
> > >
> > > I just read the slides and some of your code.
> > >
> > > I see that there may be some problems in your method, for example, there
> > > is a long HTTP request which is sent through two packets, how can it be
> > > handled in your method? Maybe I miss something.
> >         if two packets,the first packet will redirect by it's header. The
> > second packet or any following packets, I redirect them to the same real
> > server.
> >
>
> No, you probably didn't understand my question. You cannot redirect the
> first packet because its header (URL) is not finished, this is a long URL
> overlap packet. OK, I give you a detailed example, the requested URL
> is "GET /directory/image/picture.gif", the first packet has data "... GET
> /directory/image/picture.g" and the second one has data "if ...", how
> could you forward the first packet when the requested URL is not
> complete? It is broken. I think that the situation above is quite
> possible.
>
> Regards,
>
> Wensong

I think I need to buffer enough packets to figure out the request's
content type.  A very long URL or small packet size will cause this
problem. A Foundry's agent told us that their serveriron has a special
feature that they can solve the problem you mention above. And he told us
other products such as BigIP, they can't solve this problem. But, I had no
time to test. I don't sure that the words is right or not.

                                        Ping Tsai

>
> > >
> > > The other problem is that when there are more servers (e.g. 4), in most
> > > cases, the director cannot forward the request to the right server
> > ("guess
> > > right"), then the director send RST to the former server. It will bring
> > > more overhead to both the director and the real servers.
> >     When the pcket filter at the real server receives the RST packet, it
> > will not send the RST packet to the kernel. It is just a signal to tell
> > the
> > packet filter "the connection has moved to another real server", and the
> > real server will keep silent. But, after 10~15 sceonds, no other requests
> > come to this real server using the same connection. The pakcet filter must
> > close the connection.
> >
> >                                             Ping Tsai
> > >
> > > For a high-performance Layer-7 switching, maybe it can be constructured
> > as
> > > one IPVS and many KTCPVS (that can be running as real servers, just like
> > > that you install filter modules on real servers).
> > >
> > > Regards,
> > >
> > > Wensong
> > >
> >
> >
> > _______________________________________________
> > LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
> > Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
> > or go to http://www.in-addr.de/mailman/listinfo/lvs-users
> >
>
>
> _______________________________________________
> LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
> Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
> or go to http://www.in-addr.de/mailman/listinfo/lvs-users
>



<Prev in Thread] Current Thread [Next in Thread>