cluster

The cluster section has four attributes: name, owner, latlong and url.

For example,

  cluster {
    name = "Millennium Cluster"
    owner = "UC Berkeley CS Dept."
    latlong = "N37.37 W122.23"
    url = "http://www.millennium.berkeley.edu/"
  }

The name attributes specifies the name of the cluster of machines. The owner tag specifies the administrators of the cluster. The pair name/owner should be unique to all clusters in the world.

The latlong attribute is the latitude and longitude GPS coordinates of this cluster on earth. Specified to 1 mile accuracy with two decimal places per axis in decimal.

The url for more information on the cluster. Intended to give purpose, owner, administration, and account details for this cluster.

There directives directly control the XML output of gmond. For example, the cluster configuration example above would translate into the following XML.

  <CLUSTER NAME="Millennium Cluster" OWNER="UC Berkeley CS Dept."
           LATLONG="N37.37 W122.23" URL="http://www.millennium.berkeley.edu/">
  ...
  </CLUSTER>

host

 host {
   location = "1,2,3"
 }

The numbers represent Rack, Rank and Plane respectively.

globals

For example,

  globals {
    daemonize = true
    setuid = true
    user = ganglia
    host_dmax = 3600
    host_tmax = 40
  }

The daemonize attribute is a boolean. When true, gmond will daemonize. When false, gmond will run in the foreground.

The setuid attribute is a boolean. When true, gmond will set its effective UID to the uid of the user specified by the user attribute. When false, gmond will not change its effective user.

The debug_level is an integer value. When set to zero (0), gmond will run normally. A debug_level greater than zero will result in gmond running in the foreground and outputting debugging information. The higher the debug_level the more verbose the output.

The mute attribute is a boolean. When true, gmond will not send data regardless of any other configuration directives.

The deaf attribute is a boolean. When true, gmond will not receive data regardless of any other configuration directives.

The allow_extra_data attribute is a boolean. When false, gmond will not send out the EXTRA_ELEMENT and EXTRA_DATA parts of the XML. This might be useful if you are using your own frontend to the metric data and will like to save some bandwith.

The host_dmax value is an integer with units in seconds. When set to zero (0), gmond will never delete a host from its list even when a remote host has stopped reporting. If host_dmax is set to a positive number then gmond will flush a host after it has not heard from it for host_dmax seconds. By the way, dmax means "delete max".

The host_tmax value is an integer with units in seconds. This value represents the maximum amount of time that gmond should wait between updates from a host. As messages may get lost in the network, gmond will consider the host as being down if it has not received any messages from it after 4 times this value. For example, if host_tmax is set to 20, the host will appear as down after 80 seconds with no messages from it. By the way, tmax means "timeout max".

The cleanup_threshold is the minimum amount of time before gmond will cleanup any hosts or metrics where tn > dmax a.k.a. expired data.

The gexec boolean allows you to specify whether gmond will announce the hosts availability to run gexec jobs. Note: this requires that gexecd is running on the host and the proper keys have been installed.

The send_metadata_interval establishes an interval in which gmond will send or resend the metadata packets that describe each enabled metric. This directive by default is set to 0 which means that gmond will only send the metadata packets at startup and upon request from other gmond nodes running remotely. If a new machine running gmond is added to a cluster, it needs to announce itself and inform all other nodes of the metrics that it currently supports. In multicast mode, this isn't a problem because any node can request the metadata of all other nodes in the cluster. However in unicast mode, a resend interval must be established. The interval value is the minimum number of seconds between resends.

The override_hostname and override_ip parameters allow an arbitrary hostname and/or IP (hostname can be optionally specified without IP) to use when identifying metrics coming from this host.

The module_dir is an optional parameter indicating the directory where the DSO modules are to be located. If absent, the value to use is set at configure time with the --with-moduledir option which will default if omitted to the a subdirectory named "ganglia" in the directory where libganglia will be installed.

For example, in a 32-bit Intel compatible Linux host that is usually:

  /usr/lib/ganglia

udp_send_channel

The udp_send_channel has a total of seven attributes: mcast_join, mcast_if, host, port, ttl, bind and bind_hostname. bind and bind_hostname are mutually exclusive.

For example, the 2.5.x version gmond would send on the following single channel by default...

  udp_send_channel {
    mcast_join = 239.2.11.71
    port       = 8649
  }

The mcast_join and mcast_if attributes are optional. When specified gmond will create the UDP socket and join the mcast_join multicast group and send data out the interface specified by mcast_if.

You can use the bind attribute to bind to a particular local address to be used as the source for the multicast packets sent or let gmond resolve the default hostname if bind_hostname = yes.

If only a host and port are specified then gmond will send unicast UDP messages to the hosts specified.

You could specify multiple unicast hosts for redundancy as gmond will send UDP messages to all UDP channels.

Be careful though not to mix multicast and unicast attributes in the same udp_send_channel definition.

For example...

  udp_send_channel {
    host = host.foo.com
    port = 2389
  }
  udp_send_channel {
    host = 192.168.3.4
    port = 2344
  }

would configure gmond to send messages to two hosts. The host specification can be an IPv4/IPv6 address or a resolvable hostname.

The ttl attribute lets you modify the Time-To-Live (TTL) of outgoing messages (unicast or multicast).

udp_recv_channel

The udp_recv_channel section has following attributes: mcast_join, bind, port, mcast_if, family, retry_bind and buffer. The udp_recv_channel can also have an acl definition (see ACCESS CONTROL LISTS below).

For example, the 2.5.x gmond ran with a single udp receive channel...

  udp_recv_channel {
    mcast_join = 239.2.11.71
    bind       = 239.2.11.71
    port       = 8649
  }

The mcast_join and mcast_if should only be used if you want to have this UDP channel receive multicast packets the multicast group mcast_join on interface mcast_if. If you do not specify multicast attributes then gmond will simply create a UDP server on the specified port.

You can use the bind attribute to bind to a particular local address.

The family address is set to inet4 by default. If you want to bind the port to an inet6 port, you need to specify that in the family attribute. Ganglia will not allow IPV6=>IPV4 mapping (for portability and security reasons). If you want to listen on both inet4 and inet6 for a particular port, explicitly state it with the following:

  udp_recv_channel {
    port = 8666
    family = inet4
  }
  udp_recv_channel {
    port = 8666
    family = inet6
  }

If you specify a bind address, the family of that address takes precedence. f your IPv6 stack doesn't support IPV6_V6ONLY, a warning will be issued but gmond will continue working (this should rarely happen).

Multicast Note: for multicast, specifying a bind address with the same value used for mcast_join will prevent unicast UDP messages to the same port from being processed.

The sFlow protocol (see http://www.sflow.org) can be used to collect a standard set of performance metrics from servers. For servers that don't include embedded sFlow agents, an open source sFlow agent is available on SourceForge (see http://host-sflow.sourceforge.net).

To configure gmond to receive sFlow datagrams, simply add a udp_recv_channel with the port set to 6343 (the IANA registered port for sFlow):

  udp_recv_channel {
    port = 6343
  }

Note: sFlow is unicast protocol, so don't include mcast_join join. Note: To use some other port for sFlow, set it here and then specify the port in an sflow section (see below).

gmond will fail to run if it can't bind to all defined udp_recv_channels. Sometimes, on machines configured by DHCP, for example, the gmond daemon starts before a network address is assigned to the interface. Consequently, the bind fails and the gmond daemon does not run. To assist in this situation, the boolean parameter retry_bind can be set to the value true and then the daemon will not abort on failure, it will enter a loop and repeat the bind attempt every 60 seconds:

  udp_recv_channel {
    port = 6343
    retry_bind = true
  }

If you have a large system with lots of metrics, you might experience UDP drops. This happens when gmond is not able to process the UDP fast enough from the network. In this case you might consider changing your setup into a more distributed setup using aggregator gmond hosts. Alternatively you can choose to create a bigger receive buffer:

  udp_recv_channel {
    port = 6343
    buffer = 10485760
  }
B<buffer> is specified in bytes, i.e.: 10485760 will allow 10MB UDP 
to be buffered in memory.

Note: increasing buffer size will increase memory usage by gmond

tcp_accept_channel

The tcp_accept_channel has the following attributes: bind, port, interface, family and timeout. A tcp_accept_channel may also have an acl section specified (see ACCESS CONTROL LISTS below).

For example, 2.5.x gmond would accept connections on a single TCP channel.

  tcp_accept_channel {
    port = 8649
  }

The bind address is optional and allows you to specify which local address gmond will bind to for this channel.

The port is an integer than specifies which port to answer requests for data.

The family address is set to inet4 by default. If you want to bind the port to an inet6 port, you need to specify that in the family attribute. Ganglia will not allow IPV6=>IPV4 mapping (for portability and security reasons). If you want to listen on both inet4 and inet6 for a particular port, explicitly state it with the following:

  tcp_accept_channel {
    port = 8666
    family = inet4
  }
  tcp_accept_channel {
    port = 8666
    family = inet6
  }

If you specify a bind address, the family of that address takes precedence. If your IPv6 stack doesn't support IPV6_V6ONLY, a warning will be issued but gmond will continue working (this should rarely happen).

The timeout attribute allows you to specify how many microseconds to block before closing a connection to a client. The default is set to -1 (blocking IO) and will never abort a connection regardless of how slow the client is in fetching the report data.

The interface is not implemented at this time (use bind).

collection_group

The metric section has the following attributes: (one of name or name_match; name_match is only permitted if pcre support is compiled in), value_threshold and title. For a list of available metric names, run the following command:

  % gmond -m

Here is an example of a collection group for a static metric...

  collection_group {
    collect_once   = yes
    time_threshold = 1800
    metric {
     name = "cpu_num"
     title = "Number of CPUs"
    }
  }

This collection_group entry would cause gmond to collect the cpu_num metric once at startup (since the number of CPUs will not change between reboots). The metric cpu_num would be send every 1/2 hour (1800 seconds). The default value for the time_threshold is 3600 seconds if no time_threshold is specified.

The time_threshold is the maximum amount of time that can pass before gmond sends all metrics specified in the collection_group to all configured udp_send_channels. A metric may be sent before this time_threshold is met if during collection the value surpasses the value_threshold (explained below).

Here is an example of a collection group for a volatile metric...

  collection_group {
    collect_every = 60
    time_threshold = 300
    metric {
      name = "cpu_user"
      value_threshold = 5.0
      title = "CPU User"
    }
    metric {
      name = "cpu_idle"
      value_threshold = 10.0
      title = "CPU Idle"
    }
  }

This collection group would collect the cpu_user and cpu_idle metrics every 60 seconds (specified in collect_every). If cpu_user varies by 5.0% or cpu_idle varies by 10.0%, then the entire collection_group is sent. If no value_threshold is triggered within time_threshold seconds (in this case 300), the entire collection_group is sent.

Each time the metric value is collected the new value is compared with the old value collected. If the difference between the last value and the current value is greater than the value_threshold, the entire collection group is send to the udp_send_channels defined.

It's important to note that all metrics in a collection group are sent even when only a single value_threshold is surpassed.

In addition a user friendly title can be substituted for the metric name by including a title within the metric section.

By using the name_match parameter instead of name, it is possible to use a single definition to configure multiple metrics that match a regular expression. The perl compatible regular expression (pcre) syntax is used. This approach is particularly useful for a series of metrics that may vary in number between reboots (e.g. metric names that are generated for each individual NIC or CPU core).

Here is an example of using the name_match directive to enable the multicpu metrics:

  metric {
    name_match = "multicpu_([a-z]+)([0-9]+)"
    value_threshold = 1.0
    title = "CPU-\\2 \\1"
  }

Note that in the example above, there are two matches: the alphabetical match matches the variations of the metric name (e.g. idle, system) while the numeric match matches the CPU core number. The second thing to note is the use of substitutions within the argument to title.

If both name and name_match are specified, then name is ignored.

Modules

   modules {
     module {
       name = "example_module"
       language = "C/C++"
       enabled = yes
       path = "modexample.so"
       params = "An extra raw parameter"
       param RandomMax {
         value = 75
       }
       param ConstantValue {
         value = 25
       }
     }
   }

sFlow

   udp_recv_channel {
     port = 7343
   }

   sflow {
     udp_port = 7343
   }

An sFlow agent running on a hypervisor may also be sending metrics for its local virtual machines. By default these metrics are ignored, but the accept_vm_metrics flag can be used to accept those metrics too, and prefix them with an identifier for each virtual machine.

   sflow {
     accept_vm_metrics = yes
   }

The sFlow feed may also contain metrics sent from HTTP or memcached servers, or from Java VMs. Extra options can be used to ignore or accept these metrics, and to indicate that there may be multiple instances per host. For example:

    sflow {
      accept_http_metrics = yes
      multiple_http_instances = yes
    }

will allow the HTTP metrics, and also mark them with a distinguishing identifier so that each instance can be trended separately. (If multiple instances are reporting and this flag is not set, the results are likely to be garbled.)

Include

include ('/etc/ganglia/conf.d/*.conf')