multipath

Multipath link aggregation (e.g. bonding) allows the simultaneous use of multiple physical links to enable increased throughput, load balancing, redundancy, and fault tolerance. There are a variety of standard policies available that can be used right out of the box with no configuration. These policies are directly inspired by the policies offered by the Linux kernel but are now offered in user-space and hence available on all platforms that ZeroTier supports (including Windows!).

Standard policies

Policy name	Fault tolerance	Min. failover (sec.)	Default Failover (sec.)	Balancing	Aggregation efficiency	Redundancy	Sequence Reordering

Policy name	Fault tolerance	Min. failover (sec.)	Default Failover (sec.)	Balancing	Aggregation efficiency	Redundancy	Sequence Reordering
`none`	None	`60+`	`60+`	none	`none`	1	No
`active-backup`	Brief interruption	`0.25`	`10`	none	`low`	1	Only during failover
`broadcast`	Fully tolerant	`N/A`	`N/A`	none	`very low`	N	Often
`balance-rr`	Self-healing	`0.25`	`10`	packet-based	`high`	1	Often
`balance-xor`	Self-healing	`0.25`	`10`	flow-based	`very high`	1	Only during failover
`balance-aware`	Self-healing	`0.25`	`10`	adaptive flow-based	`very high`	1	Only during failover and re-balance

A policy can be used easily without specifying any additional parameters:

{
  "settings":
  {
    "defaultBondingPolicy": "active-backup"
  }
}

Custom policies

To customize a bonding policy for your use-case simply specify a basePolicy and override chosen parameters. For example, to create a more aggressive active-backup policy with low monitoring overhead that will failover 0.250 seconds after it detects a link failure, one could do the following:

{
  "settings":
  {
    "defaultBondingPolicy": "aggressive-active-backup",
    "policies":
    {
      "aggressive-active-backup":
      {
        "failoverInterval": 250,
        "pathMonitorStrategy": "dynamic",
        "basePolicy": "active-backup"
      }
    }
  }
}

Manually-specifying links

Bonds are composed of multiple links. Different sets of links can be constructed for different bonding policies and used simultaneously. One can specify the links that ZeroTier should use in any given bonding policy simply by providing an array of links with names corresponding to interface names. If a user doesn't specify a set of interfaces to use, ZeroTier will assume every system interface is available for use. However, if the user does specify a set of interfaces, ZeroTier will only use what is specified. The same applies to failover rules, if none are specified, ZeroTier will failover to any operational link. On the other hand, if the user does specify failover rules and there is ever a situation where a link is available for usage but does not fit within the rules specified by the user, it will go unused.

To specify that ZeroTier should only use eth0 and eth1 as primary links, and eth2 as a backup spare and that it should prefer IPv4 over IPv6 except on eth2 where only IPv6 is allowed:

{
  "settings":
  {
    "defaultBondingPolicy": "aggressive-active-backup",
    "policies": {
      "aggressive-active-backup":
      {
        "links": {
          "eth0": {
            "ipvPref": 46,
            "failoverTo": "eth2",
            "mode": "primary"
          },
          "eth1": {
            "ipvPref": 46,
            "failoverTo": "eth2",
            "mode": "primary"
            },
          "eth2": {
            "ipvPref": 6,
            "mode": "spare"
          }
        }
      }
    }
  }
}

Additional link-specific parameters:

Peer-specific bonds

It is possible to direct ZeroTier to form a certain type of bond with specific peers of your choice:

Active backup (`active-backup`)

Traffic is sent on only one path at any given time. A different path becomes active if the current path fails. This mode provides fault tolerance with a nearly immediate fail-over. This mode does not increase total throughput.

mode: primary, spare Link option which specifies which link is the primary device. The specified device is intended to always be the active link while it is available. There are exceptions to this behavior when using different linkSelectMethod modes. There can only be one primary link in this bonding policy.
linkSelectMethod: Specifies the selection policy for the active link during failure and/or recovery events. This is similar to the Linux Kernel's primary_reselect option but with a minor extension:
- optimize: (default if user provides no failover guidance) The primary link can change periodically if a superior path is detected.
- always: (default when links are explicitly specified): Primary link regains status as active link whenever it comes back up.
- better: Primary link regains status as active link when it comes back up and (if) it is better than the currently-active link.
- failure: Primary link regains status as active link only if the currently-active link fails.

The zerotier-cli bond <peerId> rotate command will forcibly switch to the next link available from the failover queue in an active-backup bond.

Broadcast (`broadcast`)

Traffic is sent on (all) available paths simultaneously. This mode provides fault tolerance and effectively immediate failover due to transmission redundancy. This mode is a poor utilization of throughput resources and will not increase throughput but can prevent packet loss during a link failure. The only option available is dedup which will de-duplicate all packets on the receiving end if set to true.

Balance round robin (`balance-rr`)

Traffic is striped across multiple paths. Offers partial fault tolerance immediately, full fault tolerance eventually. This policy is unaware of protocols and is primarily intended for use with protocols that are not sensitive to reordering delays. The only option available for this policy is packetsPerLink which specifies the number of packets to transmit via a path before moving to the next in the RR sequence. When set to 0 a path is chosen at random for each outgoing packet. The default value is 8, low values can begin to add overhead to packet processing.

Balance XOR (`balance-xor`, similar to the Linux kernel's balance-xor with `xmit_hash_policy=layer3+4`)

Traffic is categorized into flows based on source port, destination port, and protocol type these flows are then hashed onto available links. Each flow will persist on its assigned link interface for its entire life-cycle. Traffic that does not have an assigned port (such as ICMP pings) will be randomly distributed across links. The hash function is simply: src_port ^ dst_port ^ proto.

Balance aware (`balance-aware`, similar to Linux kernel's `balance-*lb` modes)

Traffic is dynamically allocated and balanced across multiple links simultaneously according to the target allocation. Options allow for packet or flow-based processing, and active-flow reassignment. Flows mediated over a recently failed links will be reassigned in a manner that respects the target allocation of the bond. An optional balancePolicy can be specified with the following effects: flow-dynamic(default) will hash flows onto links according to target allocation and may perform periodic re-assignments in order to preserve balance. flow-static, will hash flows onto links according to target allocation but will not re-assign flows unless a failure occurs or the link is no longer operating within acceptable parameters. And lastly packet which simply load balances packets across links according to target allocation but with no concern for sequence reordering.

Available commands

The zerotier-cli bond list or listbonds command will show the current type and health status of bonds between all peers:

The zerotier-cli bond <peerId> show command prints the parameters of a bond and real-time quality metrics of its constituent links:

The zerotier-cli bond <peerId> link <action> command will provides a facility to add and remove links in real time and nominate them for usage for particular bonds:

The zerotier-cli bond <peerId> enable will forcibly enable bonding on

Logging

If you compile with ZT_TRACE=1 ZeroTier will output important bond events to stderr allowing you to see the decisions being made by the bonding layer in real-time. For instance, below is a fairly straightforward case of the creation of an active-backup bond and the subsequent failure of its primary link and instantaneous recovery on a backup link:

Link quality

ZeroTier measures various properties of a link (such as latency, throughput, jitter, packet loss ratio, etc) in order to arrive at a quality estimate. This estimate is used by bonding policies to make allocation and failover decisions:

Policy name	Role

Policy name	Role
`active-backup`	Determines the order of the failover queue. And if `activeReselect=optimize` whether a new active link is selected.
`broadcast`	Does not use quality measurements.
`balance-rr`	May trigger removal of link from bond.
`balance-xor`	May trigger removal of link from bond.
`balance-aware`	Informs flow assignments and (re-)assignments. May trigger removal of link from bond.

A link's eligibility for being included in a bond is dependent on more than perceived quality. If a path on a link begins to exhibit disruptive behavior such as extremely high packet loss, corruption, or periodic inability to process traffic it will be removed from the bond, its traffic will be appropriately reallocated and it will be punished. Punishments gradually fade and a link can be readmitted to the bond over time. However, punishments increase exponentially if applied more than once within a given window of time.

Asymmetric links

In cases where it is necessary to bond physical links that vary radically in terms of cost, throughput, latency, and or reliability, there are a couple of ways to automatically (or manually) allocate traffic among them. Traffic distribution and balancing can be either packet or flowbased. Where packet-based is suitable for protocols not susceptible to reordering penalties and flow-based is suitable for protocols such as TCP where it is desirable to keep a conversation on a single link unless we can't avoid having to re-assign it. Additionally, a target allocation of traffic used by the bonding policy can be derived/specified in the following ways:

Automatically: This is the easiest and requires no user configuration. The bonding layer measures and senses the link properties and determines a target allocation based on perceived quality and capacity. Weaker, less reliable links will have less traffic allocated to them and stronger, more reliable links will have more traffic allocated to them. Optionally, the user can specify a set of weights (totaling 1.0) to inform the bonding layer how important certain link properties are. For instance, one may primarily be concerned with latency and jitter but not total throughput:

In the absence of user guidance ZeroTier will attempt to form an understanding of each link's speed and capacity but this value can be inaccurate if the links are not routinely saturated. Therefore we provide a way to explicitly signal the capacity of each link in terms of arbitrary but relative values:

The user specifies allocation percentages (totaling 1.0). In this case quality measurements will only be used to determine a link's eligibility to be a member of a bond, now how much traffic it will carry:

Performance and overhead considerations

Only packets with internal IDs divisible by 16 are included in measurements, this amounts to about 6.25% of all traffic.
failoverInterval specifies how quickly failover should occur during a link failure. In order to accomplish this a combination of active and passive measurement techniques are employed which may result in VERB_HELLO probes being sent every failoverInterval / 4 time units. As a mitigation monitorStrategy may be set to dynamic so that probe frequency directly correlates with native application traffic.