summaryrefslogtreecommitdiffstats
path: root/README.bonding
diff options
context:
space:
mode:
Diffstat (limited to 'README.bonding')
-rw-r--r--README.bonding179
1 files changed, 179 insertions, 0 deletions
diff --git a/README.bonding b/README.bonding
new file mode 100644
index 0000000..e81b7c9
--- /dev/null
+++ b/README.bonding
@@ -0,0 +1,179 @@
+Bonding (link aggrigation)
+==========================
+
+Features
+--------
+* Full support of features offered by the bonding kernel module.
+* Selectable bonding mode using a single parameter in rc.inet1.conf.
+* Easy addition of interfaces to the bond using a parameter in rc.inet1.conf.
+* Custom bonding module options can be proided using the generic parameter
+ IFOPTS[x] in the configuration file.
+
+
+Implementation
+--------------
+[This section can be removed from the final README.bonding as it relates to
+ implimentation by Pat rather than user level configuration]
+
+Pat should add a /lib/modprobe.d/bonding.conf (preferred) or
+/etc/modprobe.d/bonding.conf file to a package (probably network-scripts), with
+content:
+ options bonding max_bonds=0
+ options rtnl-link-bond max_bonds=0
+in order to disable the automatic creation of a bond0 interface when the kernel
+module is loaded.
+
+Rationale: Even though the bond0 interface may be scheduled for creation
+because it's configured in rc.inet1.conf, having the interface in an up state
+before configuration will prevent it from being configured due to rc.inet1's
+checking of interface state before going through the process of setting up the
+interface.
+
+This check is necessary in rc.inet1 because rc.inet1 has no way of knowing
+whether the interface has been brought up by itself or by some other means -
+there is no state information stored by rc.inet1, which prevents it from
+knowing. It is also not possible to rely on the interface being configured
+with an IP address to determine if it has been configured or not - a bond
+interface may be used as part of a bridge, and would thus not have an IP
+configured to validate against.
+
+Effects: Previous versions of Slackware followed the default behaviour when
+loading the bonding module, and would automatically create a bond0 interface
+for the user to configure. This behaviour would now change to prevent a bond0
+interface being created at load time in order for rc.inet1 to be able to
+configure a user defined bond0 interface in rc.inet1.conf at a later time.
+
+Since Slackware has never had support for bond interfaces in previous versions,
+the automatic creation of a bond interface has never been a defined behaviour -
+users should not have relied on that behaviour.
+
+Effects if not used: In its present form, rc.inet1 checks whether an interface
+is up before it tries to configure it based on the interfaces present in
+rc.inet1.conf - rc.inet1 will not try to reconfigure an interface which is up
+and running, instead issuing an debug message that the interface is already up
+when being run with any of the 'start' options.
+
+Without the inclusing of the bonding.conf options file, rc.inet1 would be
+unable to work with a bond0 interface in it's present form. It is also highly
+unlikely that rc.inet1 could be coded to *not* check whether an interface is up
+before trying to configure it - this check is fundamental to the way it works.
+
+
+Configuration
+-------------
+Bonding interfaces can be configured via two new bond specific parameters in
+rc.inet1.conf, plus use of the generic IFOPTS[x] parameter. New parameters
+are:
+ BONDNICS[x]="" The space delimited list of interfaces to add to this
+ bond. The interfaces will be brought up and configured
+ while bringing up the interface, so do not need to be
+ previously defined in rc.inet1.conf. A bond can be
+ created with only 1 interface, but does not become
+ useful until at least 2 interfaces are configured.
+
+ BONDMODE[x]="" This parameter sets the bonding mode for this
+ interface. If not specified when BONDNICS[x] has been
+ used, the default is 'balance-rr'. See below for a
+ list of all bonding modes available.
+
+The following bond modes are available:
+ balance-rr This mode is also known as round-robin mode. Packets
+ are sequentially transmitted and received through each
+ interface one by one. This mode provides load
+ balancing functionality along with fault tolerance.
+ active-backup When in this mode only one interface set to active,
+ while all other interfaces are in the backup state. If
+ the active interface fails, a backup interface replaces
+ it as the only active interface in the bond. This mode
+ only provides fault tolerance, no load balancing.
+ This mode requires that the 'primary <interface>'
+ option be configured with the IFOPTS[x] parameter.
+ balance-xor The source MAC address uses exclusive or (XOR) logic
+ with the destination MAC address. This calculation
+ ensures that the same slave interface is selected for
+ each destination MAC address. This mode provides fault
+ tolerance and load balancing.
+ broadcast All packets are sent to all the slaved interfaces. This
+ mode provides fault tolerance, but may result in
+ duplicate packets arriving at the destination.
+ 802.3ad Also known as LACP. This mode creates aggregation
+ groups that share the same speed and duplex settings,
+ and it requires a switch that supports an IEEE 802.3ad.
+ This mode uses all interfaces to form the aggregation
+ group and provides fault tolerance and load balancing.
+ balance-tlb This mode ensures that the outgoing traffic
+ distribution is set according to the load on each
+ interface and that the current interface receives all
+ the incoming traffic. If the assigned interface fails
+ to receive traffic, another interface is assigned to
+ the receiving role. This provides fault tolerance and
+ load balancing.
+ balance-alb The receiving packets are load balanced through Address
+ Resolution Protocol (ARP) negotiation. This mode
+ provides fault tolerance and load balancing.
+
+Module specific interface options can be set using the the IFOPTS[x] paramter,
+which takes a pipe (|) delimited list of options for the interface. The
+following are the most useful options which can be set with IFOPTS[x]:
+ miimon Specifies the MII link monitoring frequency in milliseconds.
+ This determines how often the link state of each slaved
+ interface is checked for link failures. A value of zero
+ disables MII link monitoring, but this is NOT advised. A value
+ of 100 is a good starting point. The default value is 0, so be
+ sure to set this option with ALL modes.
+ lacp_rate This option specifies the rate at which the host will ask the
+ link partner to transmit LACPDU packets in 802.3ad mode.
+ Possible values are:
+ slow Transmit LACPDUs every 30 seconds.
+ fast Transmit LACPDUs every 1 second.
+ The default is slow, but fast is recommended.
+ primary The interface (eth0, eth2, ...) selecting which slave is the
+ primary device. The specified interface will always be the
+ active slave while it is available. Only when the primary is
+ off-line will alternate interfaces be used. This is useful
+ when one interface is preferred over another (eg, when one
+ interface has higher throughput than another). This option
+ is only valid for active-backup, balance-tlb and balance-alb
+ modes.
+ xmit_hash_policy
+ Selects the transmit hash policy to use for interface selection
+ in balance-xor, 802.3ad, and balance-tlb modes. Possible
+ values are:
+ layer2 Use XOR of source/dest hardware MAC addresses
+ and packet type ID fields to generate the hash.
+ This algorithm will place all traffic to a
+ particular network peer on the same slave.
+ layer2+3 Use a combination of layer2 and layer3 protocol
+ information (hardware MAC addresses and IP
+ addresses) to generate the hash.
+ This algorithm will place all traffic to a
+ particular network peer on the same slave.
+ This policy is intended to provide a more
+ balanced distribution of traffic than layer2
+ alone, especially in environments where a
+ layer3 gateway device is required to reach most
+ destinations.
+ layer3+4 This policy uses upper layer protocol
+ information, when available, to generate the
+ hash. This allows for traffic to a particular
+ network peer to span multiple slave interfaces,
+ although a single connection will not span
+ multiple slaves.
+ The default value is layer2. Additional (lesser used) policies
+ are available, and documented in kernel source documentation:
+ /usr/src/linux/Documentation/networking/bonding.txt
+
+The IFOPTS[x] option should always include the 'miimon' option - not using this
+option will result in network degradation.
+In 'active-backup' mode, the 'primary' option should also be supplied.
+When using '802.3ad mode', set "lacp_rate fast" for faster recovery from an
+interface failure.
+In other modes, the 'xmit_hash_policy' should be set.
+
+Full documentation of the bonding layer is available in the kernel source
+documentation: /usr/src/linux/Documentation/networking/bonding.txt.
+
+
+--
+Darren 'Tadgy' Austin.
+<darren (at) afterdark.org.uk>