MDS212 Datasheet, PDF (58/111 Pages) Zarlink Semiconductor Inc – 12-Port 10/100Mbps Ethernet Switch

MDS212
Data Sheet
16.0 Quality Of Service (QOS)
Quality of Service (QoS) provides the capability to reserve bandwidth throughout the network. This is particularly
useful for sending voice or video over the switched network. In a switched Ethernet environment, this is only
possible with Resource Reservation Protocol (RSVP), a Layer 3 protocol. In a Layer 2 switch, QoS, referred as
Class of Service (CoS) by the IEEE 802.1Q standard, provides the capability to prioritize certain tasks on the
network. This is done at the application level, where applications can set the priority when the frame is created. The
MDS212 classifying Ethernet frames according to their IEEE 802.1p/Q VLAN priorities. There are three bits in the
VLAN ID reserved to designate the priority of a packet.
Each port stores its transmission jobs into four transmission scheduling queues, one for each internal priority.
Before transmitting, a port selects a queue from which a transmission job is read. The transmission job points to a
frame stored in memory that is fetched and transmitted. The four queues, representing four classes of traffic, are
selected using a weighted round robin (WRR) strategy. The relative service rates among these queues are
programmable such that bandwidth can be allocated according to classes. This ensures that critical applications
get a fair share of bandwidth, even when the network is overloaded.
The Search Engine recognizes the IEEE 802.1p priority tag and classifies each incoming frame into four internal
priority classes: P0, P1, P2, and P3, in decreasing priority. Since the IEEE 802.1p/Q allows up to eight priorities, a
programmable mapping allows the user to map the 802.1p priority to the internal priority tag via register AVTC.
16.1 Weighted Round Robin Transmission Strategy
Frames of four different priorities are transmitted according to a weighed round robin (WRR) strategy. The WRR is
a modified form of the fair round-robin strategy, in which the server visits the queues in turn. In a fair round-robin
strategy, the server treats all queues equally and visits them with identical frequency. In a WRR, the queues are
weighted, i.e., one queue may be visited more frequently than another. These weighs are programmable via
register AXSC, in which the service rate ratio between two adjacent classes of traffic is set.
In register AXSC, setting QSW0=2, QSW1=QSW2=1 gives the service ratio 8:2:1:1, which is a good start for most
LAN switches. This ratio allocates 67% = 8/12 of bandwidth to P0, 16% = 2/12 of bandwidth to P1, and P2 and P3
each receives 8.3% = 1/12 of bandwidth, assuming all frames have identical frame length.
16.2 Buffer Management Functions
The MDS212 stores frame data in frame buffers. The number of frame buffers in a system is the maximum number
of frames a device can store. When all frame buffers are used, incoming frames cannot enter the memory and are
discarded. Without buffer management, a congested port causes a backlog of frames that eventually occupy all
frame buffers. The MDS212 features buffer management functions that prevent a single type of traffic from
depleting all frame buffers. The buffer manager limits the number of frames each destination port can store, thereby
preventing congested ports from occupying all the buffers and blocking incoming frames.
The buffer manager examines the destination port of every frame stored, and increments a counter associated with
this destination port. These buffer counters keep track of the number of buffers occupied by frames destined to
each port. If the counter reaches a threshold, incoming frames destined for the associated port will be dropped.
This threshold is programmable via register BCT and BCHL. Register BCT allows the user to program two
thresholds, one high and one low. The user specifies a threshold, high or low, for each port in register BCHL.
The buffer manager also prevents multicast frames from occupying all frame buffers. A programmable threshold,
register MBCR, limits the number of multicast frames stored in memory. In another word, buffers are reserved for
unicast frames.
A multicast forwarding job points to a multicast frame in memory fetched and forwarded by the Frame Engine
across the Xpipe to the remote device. The Frame Engine can only forward a handful of multicast frames
simultaneously across the Xpipe. Excess multicast forwarding jobs are stored in an internal FIFO, called the MC
forwarding FIFO. If the MC forwarding FIFO is full, incoming multicast frames can no longer be forwarded to the
remote device. For these blocked multicast frames, their remote destination ports are discarded.
58
Zarlink Semiconductor Inc.