LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: [lvs-users] Question on SH scheduler

To: "LinuxVirtualServer.org users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] Question on SH scheduler
Cc: Alexander Holler <holler@xxxxxxxxxxxxx>
From: Simon Horman <horms@xxxxxxxxxxxx>
Date: Tue, 28 Feb 2012 16:12:46 +0900
On Tue, Feb 28, 2012 at 11:36:11AM +0530, krishna prasad wrote:
> On Tue, Feb 28, 2012 at 11:18 AM, Alexander Holler 
> <holler@xxxxxxxxxxxxx>wrote:
> 
> > Am 28.02.2012 06:35, schrieb krishna prasad:
> >
> >> On Tue, Feb 28, 2012 at 10:44 AM, Alexander Holler<holler@xxxxxxxxxxxxx>*
> >> *wrote:
> >>
> >>  Am 28.02.2012 06:11, schrieb krishna prasad:
> >>>
> >>>  On Tue, Feb 28, 2012 at 10:04 AM, Alexander Holler<holler@xxxxxxxxxxxxx
> >>> >*
> >>>
> >>>> *wrote:
> >>>>
> >>>>
> >>>>  Hello,
> >>>>
> >>>>>
> >>>>> Am 28.02.2012 04:26, schrieb krishna prasad:
> >>>>>
> >>>>>
> >>>>>     But I strongly think that it it good to have IP+port hashing, for
> >>>>> cases
> >>>>>
> >>>>>  where multiple clients run on single host, in this case
> >>>>>> the connections have same IP but different port. In this case also the
> >>>>>> same
> >>>>>> is desirable,i.e same client to the same real-server.
> >>>>>> This may not make a real use case for web world, but a strong case for
> >>>>>> non-web deployments like in telecom.I know LVS is increasingly used in
> >>>>>> other than web services.
> >>>>>>
> >>>>>>
> >>>>>>  What should be the use case for this? Source ports are almost always
> >>>>> choosen randomly, so you woould get the same results as balancing
> >>>>> randomly.
> >>>>>
> >>>>>
> >>>>>  Not necessarily,I came across few implementations where client port is
> >>>> fixed (they bind() port while creating socket), but I agree that most of
> >>>>
> >>>>
> >>> Sure.
> >>>
> >>>
> >>>  the times source port is random.
> >>>
> >>>> The good approach would be LVS to provide options  for IP+port or just
> >>>> IP
> >>>> hashing.
> >>>>
> >>>>
> >>> But I still miss the use case. If the client always does come with the
> >>> same port, it doesn't make a difference if the port is used too for
> >>> hashing
> >>> or just the IP.
> >>>
> >>>
> >> Oh..Sorry, I should have been clear.
> >> Here is the use case: I have a client (from out side it looks like a piza
> >> box, but internally it has many CPU..something like blade server/ATCA)
> >> which initiates TCP/SCTP connections with same IP address but with
> >> different Port.
> >> So if I use SH, all these connections (potentially this client can
> >> initiate
> >> as many as 40 connections) will land on a same real server which may not
> >> be
> >> what we wanted. we wanted the connections to be balanced (based on
> >> IP+port)
> >> across all the real servers. Does it make sense?
> >>
> >
> > Hmm, if it comes to a few thousand different ports, it would, but not for
> > 40 (imho). ;)
> >
> > Anyway, this discussion should be held at the ML, not private. Maybe
> > someone else could have add some ideas.
> >
> 
> Oh..did not notice this, this time I am including the ML group.

Hi,

I think that the scenario described above does make sense and that
none of the existing schedulers cater to it. Although persistence could be
used to achieve much if not all of the desired result. I would be happy
to consider a new scheduler that implements source address+port hashing.


_______________________________________________
Please read the documentation before posting - it's available at:
http://www.linuxvirtualserver.org/

LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
or go to http://lists.graemef.net/mailman/listinfo/lvs-users

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