Cisco Nexus 9000 Series Configuration Manual page 165

Configuring Priority Flow Control
• When a no-drop class is classified based on 802.1p CoS x and assigned a internal priority value (qos-group)
of y, Cisco recommends that you use the internal priority value x to classify traffic on 802.1p CoS only,
and not on any other field. The packet priority assigned is x if the classification is not based on CoS,
which results in packets of internal priority x and y to map to the same priority x.
• The PFC feature supports up to three no-drop classes of any maximum transmission unit (MTU) size.
However, there is a limit on the number of PFC-enabled interfaces based on the following factors:
• You can define the upper limit of any MTU in the system using the systemjumbomtu command. The
MTU range is from 1500 to 9216 bytes, and the default is 9216 bytes.
• The interface QoS policy takes precedence over the system policy. PFC priority derivation also happens
in the same order.
• Ensure that you apply the same interface-level QoS policy on all PFC-enabled interfaces for both ingress
and egress.
• To achieve end-to-end lossless service over the network, Cisco recommends that you enable PFC on
each interface through which the no-drop class traffic flows (Tx/Rx).
• Cisco recommends that you change the PFC configuration when there is no traffic. Otherwise, packets
already in the Memory Management Unit (MMU) of the system might not get the expected treatment.
• Cisco recommends that you use default buffer sizes for no-drop classes or configure different input
queuing policies suitable to 10G and 40G interfaces and the no-drop class MTU size. If the buffer size
is specified through the CLI, it allocates the same buffer size for all ports irrespective of the link speed
and MTU size. Applying the same pause buffer-size on 10G and 40G interfaces is not supported.
• Do not enable WRED on a no-drop class because it results in egress queue drops.
• Dynamic load balancing cannot be enabled for internal links with PFC. You must disable DLB and enable
RTAG7 load-balancing for internal links with the port-channel load-balance internal rtag7 command.
• The dynamic load balancing (DLB) based hashing scheme is enabled by default on all internal links of
a linecard. When DLB is enabled, no-drop traffic might experience out-of-order packet delivery when
congestion on internal links occurs and PFC is applied. If applications on the system are sensitive to
out-of-order delivery, you can adjust for this by disabling DLB at the qos-group level. Disable DLB by
using the set dlb-disable action in the QoS policy-maps and the set qos-group action for no-drop classes.
In the following example assume that qos-group 1 is a no-drop class. DLB is disabled for this no-drop
class by adding the set dlb-disable action and the set qos-group action.
switch(config)# policy-map p1
switch(config-pmap-qos)# class c1
switch(config-pmap-c-qos)# set qos-group 1
switch(config-pmap-c-qos)# set dlb-disable
switch(config-pmap-c-qos)# end
• MTU size of the no-drop class
• Number of 10G and 40G ports
Caution
Irrespective of the PFC configuration, Cisco recommends that you stop traffic
before applying or removing a queuing policy that has strict priority levels at the
interface level or the system level.
Cisco Nexus 9000 Series NX-OS Quality of Service Configuration Guide, Release 7.x
Guidelines and Limitations for Priority Flow Control
151