Contents
  ========

  Introduction
  Collecting and Summarizing Host Metrics
  Answering DNS Queries
  TTLs and MXes
  Configuring lbnamed
  The Load-Balanced Client Daemon
  Duh
  Conclusion


  Introduction
  ============

  lbnamed is a load balancing name server written in Perl using the Stanford::
  DNSserver perl module.  lbnamed allows you to create dynamic groups of hosts
  that have one name in the DNS.  A host may be in multiple groups at the same
  time. For example, the name

    www.best.stanford.edu

  represents a dynamic group of 5 web servers named:

     www{1,2,3,4,5}.stanford.edu

  When someone tries to connect to www.best.stanford.edu, a DNS query is
  performed.  That query eventually gets sent to the lbnamed server which
  responds with the name of the least loaded of those 5 web servers.

  Of course advertising a web service in a subdomain like "best.stanford.edu"
  is less than desirable.  Fortunately, that can be avoided with the simple
  alias:

    www.stanford.edu -> www.best.stanford.edu

  Now when someone tries to connect to www.stanford.edu and their DNS resolver
  queries for the IP address of www.stanford.edu it gets the following answer:

    www.stanford.edu       is an alias for  www.best.stanford.edu  (1)
    www.best.stanford.edu  is an alias for  www3.stanford.edu      (2)
    www3.stanford.edu      has address      171.64.10.89           (3)

  In this "alias chain" type answer, the middle link, (2), is provided by
  lbnamed; the others are provided by the normal name service serving the
  "stanford.edu" zone.

  lbnamed can return the address of the least loaded group member instead of
  its name.  In that case, the query for the IP address of www.stanford.edu
  would get the following answer:

    www.stanford.edu       is an alias for  www.best.stanford.edu  (1)
    www.best.stanford.edu  has address      171.64.10.89           (2)

  This is somewhat less "honest", DNS-wise, because a reverse lookup on the
  IP address yields the name www3.stanford.edu, not www.best.stanford.edu.
  It does however avoid the "alias chain", which some older BIND versions
  have trouble with (note that all those old BIND versions have security
  problems, so no one should be running them now!).


  Collecting and Summarizing Host Metrics
  =======================================

  The poller script queries the load-balanced client daemon, lbcd, on hosts
  participating in load-balanced name groups.  It collects the system load
  and other data from the lbcd running on each host (more on lbcd later).
  The poller uses the collected information to calculate an overall "weight"
  for each host.  The host names and weights are written to a file that the
  lbnamed name server uses to determine which host names to pass out in
  response to queries.

  The poller calculates the weights using the following algorithm:

    // constants

    WEIGHT_PER_USER = 10
    WEIGHT_PER_LOAD_UNIT = 3

    // data retrieved from lbcd by the poller

    l1   = current host load
    tot  = total number of users logged into the host
    uniq = number of unique users logged into the host

    // data provided in the lbnamed configuration file

    sf = this host's "server factor" in the range 0 to 10

    weight = WEIGHT_PER_USER      * (0.2 * tot + 0.8 * uniq) * (10 - sf)
           + WEIGHT_PER_LOAD_UNIT *          l1              *     sf

  As you can see, the "server factor" controls the relative importance of
  interactive user sessions.  A value of zero represents a purely interactive,
  user-oriented host; a value of 10 represents a server with no logins.


  Answering DNS Queries
  =====================

  The lbnamed script uses the weights calculated by the poller and the host
  "participation factors" to determine which host's name to pass out in
  response to a DNS query.  The participation factor is a configuration
  parameter that allows for unequal load sharing between hosts in a group.

  lbnamed sorts a group of hosts by (weight / participation factor) and passes
  out the name of the host with the lowest value.  A host with a participation
  factor of 0.10 would have to have a weight ten times that of a host with a
  participation factor of 1.0 before they would sort equally.  A very small
  participation factor, say 0.001, can be used to make one host a backup for a
  set of other hosts with the default participation factor of 1.0.

  Once lbnamed has determined which host's name to use in the DNS response,
  it sends the response and increments the host's weight.  The increment is
  calculated using the following formula:

    // constants

    WEIGHT_PER_USER = 10
    WEIGHT_PER_LOAD_UNIT = 3

    // data provided in the lbnamed configuration file

    sf = this host's "server factor" in the range 0 to 10

    increment = WEIGHT_PER_USER      * (10 - sf)
              + WEIGHT_PER_LOAD_UNIT *    sf

  Note that eventually all the hosts in a group will have the same weight and
  their names will be passed out in round-robin fashion from then on.


  TTLs and MXes
  =============

  By default, lbnamed uses TTL of 0 in its DNS responses.  That tells the DNS
  not to cache the response and to always make a new query for the name in
  question.  This gives the most accurate load sharing across the hosts in a
  group.  But it may generate too many DNS queries.  If that's the case, you
  can provide a non-zero TTL value for the group in the configuration file.
  It's still best to keep the value small to promote a truly balanced load.

  DNS queries for MX records must be handled in lbnamed for groups not using
  the alias chain response type.  For those using the alias chain, the MX
  response is "inherited" from the host's real name.  For example, an MX
  query on www.stanford.edu might get the following alias chain answer:

    www.stanford.edu       is an alias for     www.best.stanford.edu   (1)
    www.best.stanford.edu  is an alias for     www3.stanford.edu       (2)
    www3.stanford.edu      mail is handled by  10 leland.stanford.edu  (3)

  As mentioned before, only link (2) comes from lbnamed, so it doesn't have to
  handle the MX data.  But without the alias chain, the answer would be:

    www.stanford.edu       is an alias for     www.best.stanford.edu   (1)
    www.best.stanford.edu  mail is handled by  10 leland.stanford.edu  (2)

  Once again part (2) is provided by the load-balanced name server, but this
  time it's the MX record, so lbnamed must have the data to respond properly.
  In order to support this, the lbnamed configuration file includes a section
  where you may provide MX information.


  Configuring lbnamed
  ===================

  lbnamed is configured using a configuration file.  The first section of the
  file lists the names of the hosts and the names of the groups in which they
  participate (remember that a group is a load-balanced name).  Each host also
  has a server factor and a participation factor for each group it's in.  The
  file may have an optional second section for listing load-balanced name TTLs
  and MXes.  Here's a short sample configuration file:

    # SF = server factor;     default participation factor = 1.0;

    host                  SF  group(participation factor)
    ####################  ##  #########################################
    foo.stanford.edu       2  quux
    bar.stanford.edu      10  www
    baz.stanford.edu       5  quux www(.01)

    # default TTL = 0 seconds;       top slice - see lbnamed POD;
    # default MX  = none;

    group           TTL  top slice   MX
    ############  #####  #########   ##################
    www               6          0   mail.stanford.edu


  The Load-Balanced Client Daemon
  ===============================

  The load-balanced client daemon, lbcd, runs on the hosts participating in
  load-balanced name groups.  The poller queries lbcd to get the host load
  and other information.  lbcd is a mildly complex beast because it has to
  understand how to get all that information from a plethora of Unix
  variants.  Because of that complexity, lbcd is distributed and maintained
  separately from lbnamed.

  If lbcd doesn't support your flavor of Unix, or round-robin satisfies your
  load-balancing needs, or you just want to play with lbnamed without down-
  loading lbcd, this package does include slbcd, a static/simple load-balanced
  client daemon.  slbcd is written in perl.  It provides a complete lbcd with
  hard-coded values for the load and other data.  When you run it on the hosts
  participating in a load-balanced name group, they will always have the same
  weight and therefore their names will always be passed out in round-robin
  fashion.


  Duh
  ===

  The lbnamed and lbnamed.config files in this package contain Stanford-
  specific configuration data.  Unless you're going to run a load-balanced
  name service for Stanford, you'll want to edit those files replacing that
  data with values appropriate for your DNS domain.


  Conclusion
  ==========

  Hopefully this has been enough of an introduction to lbnamed that you can
  make it work for you.  Be sure to check out the POD documentation for both
  lbnamed and the poller, and the sample configuration file, lbnamed.config.

  Share and enjoy.