Dell SmartFabric OS10 User Guide Release 10.5.3

PDF

Low Latency Modes

Low latency describes a system network that processes a high volume of data messages with minimal delay (latency). These networks support operations that require near real-time access to rapidly changing data. Use Low Latency mode to reduce the switching latency for timing-critical applications such as storage networks. By default, Low Latency mode is not enabled in OS10 switches. To achieve low latency, only the Memory Management Unit (MMU) Cut-Through (CT) mode is enabled.

Low Latency modes include bypass paths in different blocks within the ingress and egress switching pipeline.
MMU CT mode switches send the packets to the destination port without buffering the entire packet in the MMU buffer.

Low latency modes

Cut-through switching mode

CT switching offers low-latency performance for SCSI traffic. Use CT switching in packet-switching systems. The switch forwards packets or frames to its destination immediately after the destination address is processed without waiting to receive the entire data.

The egress scheduler block in the NPU pipeline schedules the packet to transmit out after the first cells of packet arrive. However, egress scheduler falls back to Store and Forward (SF) mode if the conditions are not met for CT transmission, even though you configure the switch in CT mode.

The following conditions must be met for the switch to operate in CT mode. If these conditions are not met, the switch stays in SF mode, irrespective of the configured value. For Multicast packets, all the destination ports must satisfy the following conditions:

The source and destination port speed must be within the configured range. The range can be same speed ports or fast-to-slow speed ports. For example 10G to 10G, 40G to 40G, 40G to 10G, 400G to 100G. For more information, see Restrictions and Limitations.
The destination port must not experience a back-pressure due to Priority Flow Control (PFC) or pause frames.
Do not overlap the destination port. Multiple ingress ports must not send packets to the same destination port. Similarly, one ingress port must not send multiple copies of a packet to the same port; for example, unicast or mirror copy.
The queue of the destination port must not have packets waiting for transmission in SF mode.
If the port-max shaper is configured on the egress port, outgoing packets rate on the egress port must be in the configured range. The transmitted packet rate is less than the configured maximum peak rate.
If the policer configuration is enabled on the ingress port in CT Switching mode and if the incoming packet rate is higher than the peak information rate, the excess traffic drops at ingress. The outgoing traffic which abides in the policer configuration transmits in CT mode.
If the source-to-destination port path of the packet is in CT mode, no other source ports can queue their packets to the same destination port.
CT mode switching is not allowed to the CPU port.
CT mode is not allowed to the Loopback port.
CT mode is not allowed to the management port.
Destination ports must have enough Egress Pipeline (EP) credits. Depending on the port speed, a different number of EP credits are required.
Any CPU-generated packets like STP, LLDP, IGMP, and so on that goes out of the destination port may affect the CT switching of data packets momentarily on that port.

Restrictions and limitations

When the port is operating in CT mode, you can observe the following restrictions, depending on the configuration or timing of the incoming packet, PFC message, or port speed configurations.

Layer 2/Layer 1/Layer 0, and queue level maximum shaper configurations are not considered. For example, if a queue within a port has a shaper that is configured to a rate that is less than the rate of the port, the queue can violate the shaper limit and can grow to the port rate, if the traffic rate is out of the profile.

The following table lists the CT mode support on existing OS10 platforms:

Table 1. Cut-Through mode supportCT mode support on OS10 platforms
Platform	CT Mode Support
Z9100-ON	No
S4048-ON	Yes
S6000-ON	Yes (only 40G ports)
S4000-ON	Yes (only 40G ports)
S6010-ON	Yes (only 40G ports)
S4100-ON Series	Yes (only 40G ports)
Z9264-ON	Yes
S4200-ON	No
S5148-ON	No
S5200-ON Series	Yes
S3048-ON	No
Z9300-ON Series	Yes
N-Series	No

The destination port cannot use Link Level Flow Control (LLFC) Pause as flow control (PFC). However, PFC can co-exist with CT mode, but the response time can be delayed depending on the size of the on-the-fly packet.
CT mode is not allowed for multicast packets. This limitation is specific to Z9300-ON Series switches.
CT mode is not supported for egress encapsulations such as VXLAN, ERSPAN, and any other tunneling the knowledge of the full packet length.

Fast-to-slow CT mode support is available only when the maximum port speed ratio is 4:1; for example, 40G to 10G or 400G to 100G. Fast-to-slow CT mode support is available on S5200-ON Series, Z9200-ON Series, and Z9300-ON Series switches.

Table 2. Fast-to-slow CT mode supportIngress and egress port speeds
Ingress Port speed range (Min-Max)	Egress Port speed	Combinations (SRC-DEST)
10G-50G	10G	(25G-10G), (40G-10G), (50G-10G)
25G-50G	25G	(40G-25G), (50G-25G)
40G-100G	40G	(50G-40G), (100G-40G)
50G-100G	50G	(100G-50G)