Understanding Junos CoS Components

Code-Point Aliases

A code-point alias assigns a name to a pattern of code-point bits. You can use this name instead of the bit pattern when you configure other CoS components such as classifiers and rewrite rules.

Policers

Policers limit traffic of a certain class to a specified bandwidth and burst size. Packets exceeding the policer limits can be discarded, or can be assigned to a different forwarding class, a different loss priority, or both. You define policers with filters that you can associate with input interfaces.

Classifiers

Packet classification associates incoming packets with a particular CoS servicing level. In Junos OS, classifiers associate packets with a forwarding class and loss priority and assign packets to output queues based on the associated forwarding class. Junos OS supports two general types of classifiers:

• Behavior aggregate (BA) or CoS value traffic classifiers—Examine the CoS value in the packet header. The value in this single field determines the CoS settings applied to the packet. BA classifiers allow you to set the forwarding class and loss priority of a packet based on the Differentiated Services code point (DSCP) value, IEEE 802.1p value, or MPLS EXP value.

• Multifield traffic classifiers—Examine multiple fields in the packet, such as source and destination addresses and source and destination port numbers of the packet. With multifield classifiers, you set the forwarding class and loss priority of a packet based on firewall filter rules.

On switches that require the separation of unicast and multidestination (multicast, broadcast, and destination lookup fail) traffic, you create separate unicast classifiers and multidestination classifiers. You cannot assign unicast traffic and multidestination traffic to the same classifier. You can apply unicast classifiers to one or more interfaces. Multidestination classifiers apply to all of the switch interfaces and cannot be applied to individual interfaces. Switches that require the separation of unicast and multidestination traffic have 12 output queues to provide 4 output queues reserved for multidestination traffic.

On switches that do not separate unicast and multidestination traffic, unicast and multidestination traffic use the same classifiers, and you do not create a separate special classifier for multidestination traffic. Switches that do not separate unicast and multidestination traffic have eight output queues because no extra queues are required to separate the traffic.

Forwarding Classes

Forwarding classes group packets for transmission and CoS. You assign each packet to an output queue based on the packet’s forwarding class. Forwarding classes affect the forwarding, scheduling, and rewrite marking policies applied to packets as they transit the switch.

Switches provide up to five default forwarding classes:

• best-effort—Best-effort traffic

• fcoe—Fibre Channel over Ethernet traffic

• no-loss—Lossless traffic

• network-control—Network control traffic

• mcast—Multicast traffic

Note:

The default mcast forwarding class applies only to switches that require the separation of unicast and multidestination (multicast, broadcast, and destination lookup fail) traffic. On these switches, you create separate forwarding classes for the two types of traffic. The default mcast forwarding class transports only multidestination traffic, and the default best-effort, fcoe, no-loss, and network-control forwarding classes transport only unicast traffic. Unicast forwarding classes map to unicast output queues, and multidestination forwarding classes map to multidestination output queues. You cannot assign unicast traffic and multidestination traffic to the same forwarding class or to the same output queue. Switches that require the separation of unicast and multidestination traffic have 12 output queues, 8 for unicast traffic and 4 for multidestination traffic.

On switches that do not separate unicast and multidestination traffic, unicast and multidestination traffic use the same forwarding classes and output queues, so the mcast forwarding class is not valid. You do not create separate forwarding classes for multidestination traffic. Switches that do not separate unicast and multidestination traffic have eight output queues because no extra queues are required to separate the traffic.

Switches support a total of either 12 forwarding classes (8 unicast forwarding classes and 4 multicast forwarding classes), or 8 forwarding classes (unicast and multidestination traffic use the same forwarding classes), which provides flexibility in classifying traffic.

Forwarding Class Sets

You can group forwarding classes (output queues) into forwarding class sets to apply CoS to groups of traffic that require similar treatment. Forwarding class sets map traffic into priority groups to support enhanced transmission selection (ETS), which is described in IEEE 802.1Qaz.

You can configure up to three unicast forwarding class sets and one multicast forwarding class set. For example, you can configure different forwarding class sets to apply CoS to unicast groups of local area network (LAN) traffic, storage area network (SAN) traffic, and high-performance computing (HPC) traffic, and configure another group for multicast traffic.

Within each forwarding class set, you can configure special CoS treatment for the traffic mapped to each individual queue. This provides the ability to configure CoS in a two-tier hierarchical manner. At the forwarding class set tier, you configure CoS for groups of traffic using a traffic control profile. At the queue tier, you configure CoS for individual output queues within a forwarding class set using a scheduler that you map to a queue (forwarding class) using a scheduler map.

Flow Control (Ethernet PAUSE, PFC, and ECN)

Ethernet PAUSE (described in IEEE 802.3X) is a link-level flow control mechanism. During periods of network congestion, Ethernet PAUSE stops all traffic on a full-duplex Ethernet link for a period of time specified in the PAUSE message.

Note:

QFX10000 switches do not support Ethernet PAUSE.

Priority-based flow control (PFC) is described in IEEE 802.1Qbb as part of the IEEE data center bridging (DCB) specifications for creating a lossless Ethernet environment to transport loss-sensitive flows such as Fibre Channel over Ethernet (FCoE) traffic.

PFC is a link-level flow control mechanism similar to Ethernet PAUSE. However, Ethernet PAUSE stops all traffic on a link for a period of time. PFC decouples the pause function from the physical link and divides the traffic on the link into eight priorities (3-bit IEEE 802.1p code points). You can think of the eight priorities as eight “lanes” of traffic. You can apply pause selectively to the traffic on any priority without pausing the traffic on other priorities on the same link.

The granularity that PFC provides allows you to configure different levels of CoS for different types of traffic on the link. You can create lossless lanes for traffic such as FCoE, LAN backup, or management, while using standard frame-drop methods of congestion management for IP traffic on the same link.

Note:

If you transport FCoE traffic, you must enable PFC on the priority assigned to FCoE traffic (usually IEEE 802.1p code point 011 on interfaces that carry FCoE traffic).

Explicit congestion notification (ECN) enables end-to-end congestion notification between two endpoints on TCP/IP based networks. ECN must be enabled on both endpoints and on all of the intermediate devices between the endpoints for ECN to work properly. Any device in the transmission path that does not support ECN breaks the end-to-end ECN functionality. ECN notifies networks about congestion with the goal of reducing packet loss and delay by making the sending device decrease the transmission rate until the congestion clears, without dropping packets. RFC 3168, The Addition of Explicit Congestion Notification (ECN) to IP, defines ECN.

WRED Profiles and Tail Drop

A weighted random early detection (WRED) profile (drop profile) defines parameters that enable the network to drop packets during periods of congestion. A drop profile defines the conditions under which packets of different loss priorities drop, by determining the probability of dropping a packet for each loss priority when output queues become congested. Drop profiles essentially set a value for a level of queue fullness—when the queue fills to the level of the queue fullness value, packets drop. The combination of queue fill level, the probability of dropping a packet at that fill level, and loss priority of the packet, determine whether a packet is dropped or forwarded. Each pairing of a fill level with a drop probability creates a point on a drop profile curve.

You can associate different drop profiles with different loss priorities to set the probability of dropping packets. You can apply a drop profile for each loss priority to a forwarding class (output queue) by applying a drop profile to a scheduler, and then mapping the scheduler to a forwarding class using a scheduler map. When the queue mapped to the forwarding class experiences congestion, the drop profile determines the level of packet drop for traffic of each loss priority in that queue.

Loss priority affects the scheduling of a packet without affecting the packet’s relative ordering. Typically you mark packets exceeding a particular service level with a high loss priority.

Tail drop is a simple drop mechanism that drops all packets indiscriminately during periods of congestion, without differentiating among the packet loss priorities of traffic flows. Tail drop requires only one curve point that corresponds to the maximum depth of the output queue, and drop probability when traffic exceeds the buffer depth is 100 percent (all packets that cannot be stored in the queue are dropped). WRED is superior to tail-drop because WRED enables you to treat traffic of different priorities in a differentiated manner, so that higher priority traffic receives preference, and because of the ability to set multiple points on the drop curve.

Schedulers

Each switch interface has multiple queues assigned to store packets. The switch determines which queue to service based on a particular method of scheduling. This process often involves determining the sequence in which different types of packets should be transmitted.

You can define the scheduling priority (priority), minimum guaranteed bandwidth (transmit-rate), maximum bandwidth (shaping-rate), and WRED profiles to be applied to a particular queue (forwarding class) for packet transmission. By default, extra bandwidth is shared among queues in proportion to the minimum guaranteed bandwidth of each queue. On switches that support the excess-rate statement, you can configure the percentage of shared extra bandwidth an output queue receives independently from the minimum guaranteed bandwidth transmit rate, or you can use default bandwidth sharing based on the transmit rate.

A scheduler map associates a specified forwarding class with a scheduler configuration. You can associate up to four user-defined scheduler maps with the interfaces.

Rewrite Rules

A rewrite rule sets the appropriate CoS bits in the outgoing packet. This allows the next downstream device to classify the packet into the appropriate service group. Rewriting (marking) outbound packets is useful when the switch is at the border of a network and must change the CoS values to meet the policies of the targeted peer.

Note:

Ingress firewall filters can also rewrite forwarding class and loss priority values.