Hardware datapath configuration manages packet buffering, queueing and scheduling in hardware. There are two configuration input files:
/etc/cumulus/datapath/traffic.conf, which describes priority groups and assigns the scheduling algorithm and weights
/usr/lib/python2.7/dist-packages/cumulus/__chip_config/[bcm|mlx]/datapath.conf, which assigns buffer space and egress queues
While it's possible to change the buffer limits in the datapath.conf file, Cumulus Networks strongly recommends you work with a Cumulus support engineer to do so.
Each packet is assigned to an ASIC Class of Service (CoS) value based on the packet's priority value stored in the 802.1p (Class of Service) or DSCP (Differentiated Services Code Point) header field. The choice to schedule packets based on COS or DSCP is a configurable option in the
Priority groups include:
- Control: Highest priority traffic
- Service: Second-highest priority traffic
- Bulk: All remaining traffic
The scheduler is configured to use a hybrid scheduling algorithm. It applies strict priority to control traffic queues and a weighted round robin selection from the remaining queues. Unicast packets and multicast packets with the same priority value are assigned to separate queues, which are assigned equal scheduling weights.
Datapath configuration takes effect when you initialize
switchd. Changes to the
traffic.conf file require you to restart the
You can configure Quality of Service (QoS) for switches on the following platforms only:
- Broadcom Helix4, Tomahawk, Trident II, Trident II+ and Trident3
- Mellanox Spectrum
If you modify the configuration in the
/etc/cumulus/datapath/traffic.conf file, you must restart
switchd for the changes to take effect:
Example Configuration File
The following example
/etc/cumulus/datapath/traffic.conf datapath configuration file applies to 10G, 40G, and 100G switches on Broadcom Tomahawk, Trident II, Trident II+, or Trident3 and Mellanox Spectrum platforms only. However, see the note above for all the supported ASICs.
Keep in mind the following about the configuration:
- Regarding the default source packet fields and mapping, each selected packet field should have a block of mapped values. Any packet field value that is not specified in the configuration is assigned to a default internal switch priority. The configuration applies to every forwarding port unless a custom remark configuration is defined for that port (see below).
- Regarding the default remark packet fields and mapping, each selected packet field should have a block of mapped values. Any internal switch priority value that is not specified in the configuration is assigned to a default packet field value. The configuration applies to every forwarding port unless a custom remark configuration is defined for that port (see below).
- Per-port source packet fields and mapping apply to the designated set of ports.
- Per-port remark packet fields and mapping apply to the designated set of ports.
On Mellanox Spectrum switches, packet priority remark must be enabled on the ingress port. A packet received on a remark-enabled port is remarked according to the priority mapping configured on the egress port. If packet priority remark is configured the same way on every port, the default configuration example above is correct. However, per-port customized configurations require two port groups: one for the ingress ports and one for the egress ports, as below:
Configure Traffic Marking through ACL Rules
You can mark traffic for egress packets through
ip6tables rule classifications. To enable these rules, you do one of the following:
- Mark DSCP values in egress packets.
- Mark 802.1p CoS values in egress packets.
To enable traffic marking, use
cl-acltool. Add the
-p option to specify the location of the policy file. By default, if you don't include the
cl-acltool looks for the policy file in
ip6tables-based marking is supported via the following action extension:
ebtables, the setqos keyword must be in lowercase, as in:
You can specify one of the following targets for SETQOS/setqos:
|--set-cos INT||Sets the datapath resource/queuing class value. Values are defined in IEEE_P802.1p.|
|--set-dscp value||Sets the DSCP field in packet header to a value, which can be either a decimal or hex value.|
|--set-dscp-class class||Sets the DSCP field in the packet header to the value represented by the DiffServ class value. This class can be EF, BE or any of the CSxx or AFxx classes.|
--set-dscp-class, but not both.
Here are two example rules:
You can put the rule in either the mangle table or the default filter table; the mangle table and filter table are put into separate TCAM slices in the hardware.
To put the rule in the mangle table, include
-t mangle; to put the rule in the filter table, omit
Configure Priority Flow Control
Priority flow control, as defined in the IEEE 802.1Qbb standard, provides a link-level flow control mechanism that can be controlled independently for each Class of Service (CoS) with the intention to ensure no data frames are lost when congestion occurs in a bridged network.
PFC is not supported on switches with the Helix4 ASIC.
PFC is a layer 2 mechanism that prevents congestion by throttling packet transmission. When PFC is enabled for received packets on a set of switch ports, the switch detects congestion in the ingress buffer of the receiving port and signals the upstream switch to stop sending traffic. If the upstream switch has PFC enabled for packet transmission on the designated priorities, it responds to the downstream switch and stops sending those packets for a period of time.
PFC operates between two adjacent neighbor switches; it does not provide end-to-end flow control. However, when an upstream neighbor throttles packet transmission, it could build up packet congestion and propagate PFC frames further upstream: eventually the sending server could receive PFC frames and stop sending traffic for a time.
The PFC mechanism can be enabled for individual switch priorities on specific switch ports for RX and/or TX traffic. The switch port’s ingress buffer occupancy is used to measure congestion. If congestion is present, the switch transmits flow control frames to the upstream switch. Packets with priority values that do not have PFC configured are not counted during congestion detection; neither do they get throttled by the upstream switch when it receives flow control frames.
PFC congestion detection is implemented on the switch using xoff and xon threshold values for the specific ingress buffer which is used by the targeted switch priorities. When a packet enters the buffer and the buffer occupancy is above the xoff threshold, the switch transmits an Ethernet PFC frame to the upstream switch to signal packet transmission should stop. When the buffer occupancy drops below the xon threshold, the switch sends another PFC frame upstream to signal that packet transmission can resume. (PFC frames contain a quanta value to indicate a timeout value for the upstream switch: packet transmission can resume after the timer has expired, or when a PFC frame with quanta == 0 is received from the downstream switch.)
After the downstream switch has sent a PFC frame upstream, it continues to receive packets until the upstream switch receives and responds to the PFC frame. The downstream ingress buffer must be large enough to store those additional packets after the xoff threshold has been reached.
Before Cumulus Linux 3.1.1, PFC was designated as a lossless priority group. The lossless priority group has been removed from Cumulus Linux.
PFC is disabled by default in Cumulus Linux. Enabling priority flow control (PFC) requires configuring the following settings in
/etc/cumulus/datapath/traffic.conf on the switch:
- Specifying the name of the port group in
pfc.port_group_listin brackets; for example, pfc.port_group_list = [pfc_port_group].
- Assigning a CoS value to the port group in
pfc.pfc_port_group.cos_listsetting. Note that pfc_port_group is the name of a port group you specified above and is used throughout the following settings.
- Populating the port group with its member ports in
- Setting a PFC buffer size in
pfc.pfc_port_group.port_buffer_bytes. This is the maximum number of bytes allocated for storing bursts of packets, guaranteed at the ingress port. The default is 25000 bytes.
- Setting the xoff byte limit in
pfc.pfc_port_group.xoff_size. This is a threshold for the PFC buffer; when this limit is reached, an xoff transition is initiated, signaling the upstream port to stop sending traffic, during which time packets continue to arrive due to the latency of the communication. The default is 10000 bytes.
- Setting the xon delta limit in
pfc.pfc_port_group.xon_delta. This is the number of bytes to subtract from the xoff limit, which results in a second threshold at which the egress port resumes sending traffic. After the xoff limit is reached and the upstream port stops sending traffic, the buffer begins to drain. When the buffer reaches 8000 bytes (assuming default xoff and xon settings), the egress port signals that it can start receiving traffic again. The default is 2000 bytes.
- Enabling the egress port to signal the upstream port to stop sending traffic (
pfc.pfc_port_group.tx_enable). The default is true.
- Enabling the egress port to receive notifications and act on them (
pfc.pfc_port_group.rx_enable). The default is true.
- The switch priority value(s) are mapped to the specific ingress buffer for each targeted switch port. Cumulus Linux looks at either the 802.1p bits or the IP layer DSCP bits depending on which is configured in the
traffic.conffile to map packets to internal switch priority values.
The following configuration example shows PFC configured for ports swp1 through swp4 and swp6:
A port group refers to one or more sequences of contiguous ports. Multiple port groups can be defined by:
- Adding a comma-separated list of port group names to the port_group_list.
- Adding the port_set, rx_enable, and tx_enable configuration lines for each port group.
You can specify the set of ports in a port group in comma-separated sequences of contiguous ports; you can see which ports are contiguous in
/var/lib/cumulus/porttab. The syntax supports:
- A single port (swp1s0 or swp5)
- A sequence of regular swp ports (swp2-swp5)
- A sequence within a breakout swp port (swp6s0-swp6s3)
A sequence of regular and breakout ports, provided they are all in a contiguous range. For example:
switchd to allow the PFC configuration changes to take effect:
Configure Link Pause
The PAUSE frame is a flow control mechanism that halts the transmission of the transmitter for a specified period of time. A server or other network node within the data center may be receiving traffic faster than it can handle it, thus the PAUSE frame. In Cumulus Linux, individual ports can be configured to execute link pause by:
- Transmitting pause frames when its ingress buffers become congested (TX pause enable) and/or
- Responding to received pause frames (RX pause enable).
Link pause is disabled by default. Enabling link pause requires configuring settings in
/etc/cumulus/datapath/traffic.conf, similar to how you configure priority flow control. The settings are explained in that section as well.
What's the difference between link pause and priority flow control?
Priority flow control is applied to an individual priority group for a specific ingress port.
Link pause (also known as port pause or global pause) is applied to all the traffic for a specific ingress port.
Here is an example configuration that enables both types of link pause for swp1 through swp4 and swp6:
switchd to allow link pause configuration changes to take effect:
Configure Cut-through Mode and Store and Forward Switching
Cut-through mode is disabled in Cumulus Linux by default on switches with Broadcom ASICs. With cut-though mode enabled and link pause is asserted, Cumulus Linux generates a TOVR and TUFL ERROR; certain error counters increment on a given physical port.
To work around this issue, disable link pause or disable cut-through mode in
To disable link pause, comment out the
link_pause* section in
To enable store and forward switching, set
cut_through_enable to false in
Configure Explicit Congestion Notification
Explicit Congestion Notification (ECN) is defined by RFC 3168. ECN gives a Cumulus Linux switch the ability to mark a packet to signal impending congestion instead of dropping the packet outright, which is how TCP typically behaves when ECN is not enabled.
ECN is a layer 3 end-to-end congestion notification mechanism only. Packets can be marked as ECN-capable transport (ECT) by the sending server. If congestion is observed by any switch while the packet is getting forwarded, the ECT-enabled packet can be marked by the switch to indicate the congestion. The end receiver can respond to the ECN-marked packets by signaling the sending server to slow down transmission. The sending server marks a packet ECT by setting the least 2 significant bits in an IP header
DiffServ (ToS) field to 01 or 10. A packet that has the least 2 significant bits set to 00 indicates a non-ECT-enabled packet.
The ECN mechanism on a switch only marks packets to notify the end receiver. It does not take any other action or change packet handling in any way, nor does it respond to packets that have already been marked ECN by an upstream switch.
On Trident II switches only, if ECN is enabled on a specific queue, the ASIC also enables RED on the same queue. If the packet is ECT marked (the ECN bits are 01 or 10), the ECN mechanism executes as described above. However, if it is entering an ECN-enabled queue but is not ECT marked (the ECN bits are 00), then the RED mechanism uses the same threshold and probability values to decide whether to drop the packet. Packets entering a non-ECN-enabled queue do not get marked or dropped due to ECN or RED in any case.
ECN is implemented on the switch using minimum and maximum threshold values for the egress queue length. When a packet enters the queue and the average queue length is between the minimum and maximum threshold values, a configurable probability value will determine whether the packet will be marked. If the average queue length is above the maximum threshold value, the packet is always marked.
The downstream switches with ECN enabled perform the same actions as the traffic is received. If the ECN bits are set, they remain set. The only way to overwrite ECN bits is to enable it — that is, set the ECN bits to 11.
ECN is supported on Broadcom Tomahawk, Trident II, Trident II+ and Trident3, and Mellanox Spectrum switches only.
ECN is disabled by default in Cumulus Linux. You can enable ECN for individual switch priorities on specific switch ports. ECN requires configuring the following settings in
/etc/cumulus/datapath/traffic.conf on the switch:
- Specifying the name of the port group in
ecn.port_group_listin brackets; for example,
ecn.port_group_list = [ecn_port_group].
- Assigning a CoS value to the port group in
ecn.ecn_port_group.cos_list. If the CoS value of a packet matches the value of this setting, then ECN is applied. Note that ecn_port_group is the name of a port group you specified above.
- Populating the port group with its member ports (
ecn.ecn_port_group.port_set), where ecn_port_group is the name of the port group you specified above. Congestion is measured on the egress port queue for the ports listed here, using the average queue length: if congestion is present, a packet entering the queue may be marked to indicate that congestion was observed. Marking a packet involves setting the least 2 significant bits in the IP header DiffServ (ToS) field to 11.
- The switch priority value(s) are mapped to specific egress queues for the target switch ports.
ecn.ecn_port_group.probabilityvalue indicates the probability of a packet being marked if congestion is experienced.
The following configuration example shows ECN configured for ports swp1 through swp4 and swp6:
switchd to allow the ECN configuration changes to take effect:
Check Interface Buffer Status
On Mellanox switches, you can collect a fine-grained history of queue lengths using histograms maintained by the ASIC; see the ASIC monitoring chapter for details.
On Broadcom switches, the buffer status is not visible currently.