bonding: Balance bond slaves based on ratio.

[ovs.git] / vswitchd / INTERNALS

                       ========================
                        ovs-vswitchd Internals
                       ========================

This document describes some of the internals of the ovs-vswitchd
process.  It is not complete.  It tends to be updated on demand, so if
you have questions about the vswitchd implementation, ask them and
perhaps we'll add some appropriate documentation here.

Most of the ovs-vswitchd implementation is in vswitchd/bridge.c, so
code references below should be assumed to refer to that file except
as otherwise specified.

Bonding
=======

Bonding allows two or more interfaces (the "slaves") to share network
traffic.  From a high-level point of view, bonded interfaces act like
a single port, but they have the bandwidth of multiple network
devices, e.g. two 1 GB physical interfaces act like a single 2 GB
interface.  Bonds also increase robustness: the bonded port does not
go down as long as at least one of its slaves is up.

In vswitchd, a bond always has at least two slaves (and may have
more).  If a configuration error, etc. would cause a bond to have only
one slave, the port becomes an ordinary port, not a bonded port, and
none of the special features of bonded ports described in this section
apply.

There are many forms of bonding, but ovs-vswitchd currently implements
only a single kind, called "source load balancing" or SLB bonding.
SLB bonding divides traffic among the slaves based on the Ethernet
source address.  This is useful only if the traffic over the bond has
multiple Ethernet source addresses, for example if network traffic
from multiple VMs are multiplexed over the bond.

Enabling and Disabling Slaves
-----------------------------

When a bond is created, a slave is initially enabled or disabled based
on whether carrier is detected on the NIC (see iface_create()).  After
that, a slave is disabled if its carrier goes down for a period of
time longer than the downdelay, and it is enabled if carrier comes up
for longer than the updelay (see bond_link_status_update()).  There is
one exception where the updelay is skipped: if no slaves at all are
currently enabled, then the first slave on which carrier comes up is
enabled immediately.

The updelay should be set to a time longer than the STP forwarding
delay of the physical switch to which the bond port is connected (if
STP is enabled on that switch).  Otherwise, the slave will be enabled,
and load may be shifted to it, before the physical switch starts
forwarding packets on that port, which can cause some data to be
"blackholed" for a time.  The exception for a single enabled slave
does not cause any problem in this regard because when no slaves are
enabled all output packets are blackholed anyway.

When a slave becomes disabled, the vswitch immediately chooses a new
output port for traffic that was destined for that slave (see
bond_enable_slave()).  It also sends a "gratuitous learning packet" on
the bond port (on the newly chosen slave) for each MAC address that
the vswitch has learned on a port other than the bond (see
bond_send_learning_packets()), to teach the physical switch that the
new slave should be used in place of the one that is now disabled.
(This behavior probably makes sense only for a vswitch that has only
one port (the bond) connected to a physical switch; vswitchd should
probably provide a way to disable or configure it in other scenarios.)

Bond Packet Input
-----------------

Bond packet input processing takes place in process_flow().

Bonding accepts unicast packets on any bond slave.  This can
occasionally cause packet duplication for the first few packets sent
to a given MAC, if the physical switch attached to the bond is
flooding packets to that MAC because it has not yet learned the
correct slave for that MAC.

Bonding only accepts multicast (and broadcast) packets on a single
bond slave (the "active slave") at any given time.  Multicast packets
received on other slaves are dropped.  Otherwise, every multicast
packet would be duplicated, once for every bond slave, because the
physical switch attached to the bond will flood those packets.

Bonding also drops received packets when the vswitch has learned that
the packet's MAC is on a port other than the bond port itself.  This is
because it is likely that the vswitch itself sent the packet out the
bond port on a different slave and is now receiving the packet back.
This occurs when the packet is multicast or the physical switch has not
yet learned the MAC and is flooding it.  However, the vswitch makes an
exception to this rule for broadcast ARP replies, which indicate that
the MAC has moved to another switch, probably due to VM migration.
(ARP replies are normally unicast, so this exception does not match
normal ARP replies.  It will match the learning packets sent on bond
fail-over.)

The active slave is simply the first slave to be enabled after the
bond is created (see bond_choose_active_iface()).  If the active slave
is disabled, then a new active slave is chosen among the slaves that
remain active.  Currently due to the way that configuration works,
this tends to be the remaining slave whose interface name is first
alphabetically, but this is by no means guaranteed.

Bond Packet Output
------------------

When a packet is sent out a bond port, the bond slave actually used is
selected based on the packet's source MAC (see choose_output_iface()).
In particular, the source MAC is hashed into one of 256 values, and
that value is looked up in a hash table (the "bond hash") kept in the
"bond_hash" member of struct port.  The hash table entry identifies a
bond slave.  If no bond slave has yet been chosen for that hash table
entry, vswitchd chooses one arbitrarily.

Every 10 seconds, vswitchd rebalances the bond slaves (see
bond_rebalance_port()).  To rebalance, vswitchd examines the
statistics for the number of bytes transmitted by each slave over
approximately the past minute, with data sent more recently weighted
more heavily than data sent less recently.  It considers each of the
slaves in order from most-loaded to least-loaded.  If highly loaded
slave H is significantly more heavily loaded than the least-loaded
slave L, and slave H carries at least two hashes, then vswitchd shifts
one of H's hashes to L.  However, vswitchd will only shift a hash from
H to L if it will decrease the ratio of the load between H and L by at
least 0.1.

Currently, "significantly more loaded" means that H must carry at
least 1 Mbps more traffic, and that traffic must be at least 3%
greater than L's.