5
•
mode=6, adaptive load balancing (balance-alb)
—Does not require switches. This mode
integrates the balance-tlb mode and load balancing of IPv4 packet receiving. It is realized by
ARP negotiation. The bonding driver intercepts the ARP replies sent by the local device and
changes the source MAC address into a unique MAC address of a backup device in bonding,
allowing different peers to communicate with different MAC addresses. This mode is used
commonly.
Teaming
Teaming binds multiple physical network adapters on the same server into a virtual NIC through
software. For the external network, the server has only one visible NIC. For applications and the
network where the server resides, the server has only one network link or only one IP address that
can be accessed.
Teaming is supported in both Windows and Linux operating systems. This section uses the Windows
Server operating system as an example.
Teaming has the following modes:
•
Static teaming
—A switch-dependent mode in which member NICs must connect to the same
physical switch.
•
Switch independent
—Member NICs can be connected to different switches in active/standby
mode. Load balancing aggregation can be realized only when the member NICs connect to the
same switch.
•
LACP
—You must enable LACP on the switch first. This mode integrates multiple NICs into one
logical link. Data is transmitted at the fastest speed in LACP mode.
After teaming finishes, you must configure the load balancing mode. Load balancing has the
following modes:
•
Address hash mode
—In this mode, when a packet arrives at the team, the device uses the
hash algorithm to calculate the packet sending physical NIC based on the destination address
information (MAC address, IP address, and port number). This mode cannot control traffic
direction. If a large amount of traffic goes to the same destination address, the traffic will be sent
by the same physical NIC.
•
Hyper-V port mode
—Uses the MAC addresses of VMs or Hyper-V ports to which VMs are
connected as the basis for traffic distribution. Because each VM has an independent MAC
address, this mode has higher traffic distribution efficiency if compared with the address hash
mode. In this mode, data are transmitted by different physical NICs bound to the vNIC and the
binding is based on vNICs instead of VMs. As a best practice, enable this mode when you use
a Hyper-V external virtual switch.
•
Dynamic mode
—Introduced for Windows Server 2016 and later. In this mode, data is evenly
distributed to all NICs to make full use of bandwidth resources. This mode is the most optimal
load balancing mode.
TCP offloading
TCP offloading is a TCP acceleration technology. On a high speed Ethernet, for example, 10-GE
Ethernet, processing TCP/IP packet headers consumes great CPU resources. Using NIC hardware
to process the headers can ease the CPU burden.
Offload allocates some data processing work (for example, fragmentation and reassembly) which
should be done by the operating system to the NIC hardware to reduce CPU resource consumption
and enhance processing performance.
Features related to TCP are as follows:
•
TCP segmentation offload (TSO)
—Segments TCP packets.
•
Large segment offload (LSO)/large receive offload (LRO)
—When the sending data
exceeds the specified MTU, the operating system submits a transmission request to the NIC
only once. The NIC then automatically segments, encapsulates, and sends the data packets. If
a large number of fragments are received, LRO helps to assemble multiple fragments to a
larger one and submits the larger fragment to the operating system.
