Re: [load balancing] Cascading switches off of a Foundry switch

From: Alex Samonte (asamonteIZZATsitesmith.com)
Date: Fri Feb 16 2001 - 19:24:29 EST

  • Next message: Alex Samonte: "Re: [load balancing] Verisign and Load Balancers"

    On Thu, Feb 15, 2001 at 09:41:17PM -0500, Nimesh Vakharia wrote:
    >
    > On Thu, 15 Feb 2001, Alex Samonte wrote:
    >
    > >
    > > I'm not a big fan of DSR. DSR was a hack so they can double performance
    > > numbers (it's not a matter of throughput, it's packets per second)
    > >
    > Sure, its a matter of preference. A hack? lets call it a cool hack
    > :). One thing I like about it is it eliminates the SI as being the
    > bottleneck for all network traffic to machines behing the SI. Especially
    > if u'r doing some heavy transfers to those boxes. U'r also eliminating
    > rewrites on every packet on the way out. (Although its all
    > ASIC but u'r still wining on bit on efficiency). Spanning tree diameter
    > tends to decreases etc...

    It's a hack because it wasn't really designed for or intended, Just about
    ANY load balancer can do DSR it's a function of the site architecture.

    Yes, you avoid doing rewrites on the way out. And with TCP traffic in and out
    is the same number of packets, hence my statement about it doubles performance

    If you are doing heavy transfers TO those boxes, DSR won't help you.
    If you are doing heavy transfers FROM those boxes, DSR will help you
    if you need more network throughput. DSR was put in place so people could
    exceed 100mb/s before most of the load balancers truely supported gigabit
    speeds.

    > > Making modifications to my web/mail servers is not something I want to do.
    > > That's why you get a load balancer. And not all servers support binding
    > > to a bound IP. Some servers you don't have source or support for, and
    > > then you are SOL.
    > I am not sure how the binding affects DSR, every app by default
    > binds to IP0 (ie all interfaces) on the box unless configured
    > otherwise. This is probably how many people use it anyways. Its really
    > not that bad to maintain loopbacks on the box. Its a one time, one
    > statement deal or in case of windows a few clicks deal!

    Right but you have to make the server respond with the VIP alias specifically
    not just any of them (solaris had early problems with this). Every app by
    default does not bind to ip0. Some are actually smart and let you
    choose. If you don't have that choice, you have to do fun stuff with the
    OS to make it say 'always respond with this IP'. Which may negate
    performance gains you might have gotten.

    > > Having the LB in the data path is really the only way to make LBs work right
    > > and if you remove half of that stream, you also lose half of the information.
    > > When L5 and above information is making more and more of a difference, DSR just
    > > makes half of it invisible.
    >
    > This is where I agree. I am curious how people are using L5
    > information for analysis from a SI. Maybe you can help? We leave
    > it upto the logging on the respective servers. Personally DSR
    > looks cleaner from a design perspective and just gives you an option
    > besides the norm. What u do depends on your design scenario and
    > priorities. Its like sometimes you win some u lose some.

    I wouldn't call it cleaner. I wouldn't really call it necessary any more either
    there are many LB options out there that can fulfill the performance
    requirements that at one time made DSR an appealing option.

    -Alex



    This archive was generated by hypermail 2b30 : Fri Feb 16 2001 - 19:29:38 EST