RE: [load balancing] Alteon and real server stats

From: P T (
Date: Fri Jan 02 2004 - 12:48:22 EST

  • Next message: Brown, Rodrick: "[load balancing] CSS Load Balancing."

    Happy new year to you as well. I always seem to enjoy the second day of the
    new year more than the first! :-)

    >>Why the sessions numbers are not carried by the Virt for each real
    >>server? eventhough the real server is used under more than 1 virt...

    The 1:5/1:8 discrepancy you are seeing is likely because of this: you are
    measuring hit counts, but the leastconns metric uses concurrent connection
    counts. For example, if you load-balanced a Starfire and a P3/500 under one
    virtual, with equal weights and 200 concurrent users total for the site
    (assume 1 connection per user for the purpose of this example), at the end
    of the day the total hit count on the Starfire would be much higher than the
    P3, even though at any given time both had roughly 100 concurrent

    Here’s one line of reasoning behind the Alteon-like design: Say you had one
    real server under 50 different virtual servers, and each virtual had other
    real servers. If each virtual based its “leastconns” decision only on the
    connections that it knew about for the shared real, the leastconns metric
    would be useless, as it would have only 1/49th of the information it needed.
    Also, if concurrent connection counts were carried by virtuals, then weights
    would also need to be applied per-virtual-per-real, but on Alteon pretty
    much everything L4-7 is defined at the real server level (it would be messy
    to make a reasonable representation of what you want given the current

    Here’s where that reasoning breaks down: leastconns (not just in Alteon, in
    any SLB) is a very useful metric, because a server’s capacity can be
    measured by its ability to not only accept, but complete (or service),
    connections. Considering that line of reasoning alone, in the example above
    each virtual would not have just 1/49th of the information needed, but 100%
    (i.e. each virtual would not need to know what the other virtuals were doing
    to that real, because as other virtuals place load on the real it simply
    begins to appear as a slower server, and less connections will be sent to
    it). Certainly there are factors other than ability to service (complete)
    connections that must be considered, such as total connection capacity of a
    server and the effect of high connection counts on performance/efficiency,
    but those can addressed by other metrics (such as weight and maxconns).

    I do agree with you that connection counts should be per-virtual-per-real.
    So should weights, URLs, health checks, etc. IMO it makes for a much more
    flexible design.


    >From: "Cihan Subasi (Garanti Teknoloji)" <>
    >To: "LoabBalancing (E-mail)" <>
    >Subject: [load balancing] Alteon and real server stats
    >Date: Fri, 2 Jan 2004 10:01:43 +0200
    >Hi all, first of all happy new year to all of you...
    >I have a real server (realA) and it has been used under two different VIRTs
    >(virtA and virtB). Under virtA I have only this real server, and under
    >virtB I have this one and another real (realB). Weight of realB is 5, and
    >realA has weight 1. But since the number of sessions for realA is the sum
    >of virtA and virtB, the real server do not get as much hit as I expect from
    >the virt that has 2 real servers. Even if it has a 1:5 ration under virtB,
    >the hit that it receives for the site VýrtB is much less than 1:5 ratio
    >(around 1:8). Why the sessions numbers are not carried by the Virt for
    >each real server? eventhough the real server is used under more than 1
    >Cihan SUBASI
    >Garanti Technology
    >Internet ve Yazilim Hizmetleri
    >Tel:(90)(212)4783426 GSM:(90)(533)(2750353)
    > <>
    >Success is a wonderful thing, but never underestimate the value of failure.
    >Failure teaches many more things than success ever can.
    >This message and attachments are confidential and intended solely for the
    >individual(s) stated in this
    >message.If you received this message although you are not the addressee you
    >are responsible to keep the
    >message confidential .The sender has no responsibility for the accuracy or
    >correctness of the
    >information in the message and its attachments.Our company shall have no
    >liability for any changes
    >or late receiving,loss of integrity and confidentiality,viruses and any
    >damages caused in
    >anyway to your computer system.
    >Bu mesaj ve ekleri mesajda gonderildigi belirtilen kisi/kisilere ozeldir ve
    >gizlidir.Bu mesajin muhatabi
    >olmamaniza ragmen tarafiniza ulasmis olmasi halinde mesaj iceriginin
    >gizliligi ve bu gizlilik yukumlulugune
    >uyulmasi zorunlulugu tarafiniz icin de soz konusudur.Mesaj ve eklerinde yer
    >alan bilgilerin dogrulugu ve
    >guncelligi konusunda gonderenin ya da sirketimizin herhangi bir sorumlulugu
    >mesajin ve bilgilerinin size degisiklige ugrayarak veya gec ulasmasindan,
    >butunlugunun ve gizliliginin
    >korunamamasindan, virus icermesinden ve bilgisayar sisteminize verebilecegi
    >herhangi bir zarardan
    >sorumlu tutulamaz.

    Have fun customizing MSN Messenger — learn how here!

    The Load Balancing Mailing List
    MRTG with SLB:
    Hosted by:

    This archive was generated by hypermail 2.1.4 : Fri Jan 02 2004 - 13:05:52 EST