Priority-Based Flow Control - Dell S6100 Configuration Manual

For example, instead of deploying an Ethernet network for LAN traffic, include additional storage area networks (SANs) to ensure lossless
Fibre Channel traffic, and a separate InfiniBand network for high-performance inter-processor computing within server clusters, only one
DCB-enabled network is required in a data center. The Dell Networking switches that support a unified fabric and consolidate multiple
network infrastructures use a single input/output (I/O) device called a converged network adapter (CNA).
A CNA is a computer input/output device that combines the functionality of a host bus adapter (HBA) with a network interface controller
(NIC). Multiple adapters on different devices for several traffic types are no longer required.
Data center bridging satisfies the needs of the following types of data center traffic in a unified fabric:
Traffic Description
LAN traffic LAN traffic consists of many flows that are insensitive to latency requirements, while certain applications, such as
streaming video, are more sensitive to latency. Ethernet functions as a best-effort network that may drop packets
in the case of network congestion. IP networks rely on transport protocols (for example, TCP) for reliable data
transmission with the associated cost of greater processing overhead and performance impact LAN traffic consists
of a large number of flows that are generally insensitive to latency requirements, while certain applications, such as
streaming video, are more sensitive to latency. Ethernet functions as a best-effort network that may drop packets
in case of network congestion. IP networks rely on transport protocols (for example, TCP) for reliable data
transmission with the associated cost of greater processing overhead and performance impact.
Storage traffic Storage traffic based on Fibre Channel media uses the Small Computer System Interface (SCSI) protocol for data
transfer. This traffic typically consists of large data packets with a payload of 2K bytes that cannot recover from
frame loss. To successfully transport storage traffic, data center Ethernet must provide no-drop service with
lossless links.
InterProcess InterProcess Communication (IPC) traffic within high-performance computing clusters to share information. Server
Communication traffic is extremely sensitive to latency requirements.
(IPC) traffic
To ensure lossless delivery and latency-sensitive scheduling of storage and service traffic and I/O convergence of LAN, storage, and server
traffic over a unified fabric, IEEE data center bridging adds the following extensions to a classical Ethernet network:
• 802.1Qbb — Priority-based Flow Control (PFC)
• 802.1Qaz — Enhanced Transmission Selection (ETS)
• 802.1Qau — Congestion Notification
• Data Center Bridging Exchange (DCBx) protocol
NOTE: Dell Networking OS supports only the PFC, ETS, and DCBx features in data center bridging.

Priority-Based Flow Control

In a data center network, priority-based flow control (PFC) manages large bursts of one traffic type in multiprotocol links so that it does not
affect other traffic types and no frames are lost due to congestion.
When PFC detects congestion on a queue for a specified priority, it sends a pause frame for the 802.1p priority traffic to the transmitting
device. In this way, PFC ensures that PFC-enabled priority traffic is not dropped by the switch.
PFC enhances the existing 802.3x pause and 802.1p priority capabilities to enable flow control based on 802.1p priorities (classes of
service). Instead of stopping all traffic on a link (as performed by the traditional Ethernet pause mechanism), PFC pauses traffic on a link
according to the 802.1p priority set on a traffic type. You can create lossless flows for storage and server traffic while allowing for loss in
case of LAN traffic congestion on the same physical interface.
The following illustration shows how PFC handles traffic congestion by pausing the transmission of incoming traffic with dot1p priority 4.
Data Center Bridging (DCB) 247