Zte ZXR10 5900 Product Description page 42

ZXR10 5900 5200 Product Description
As shown in Figure 8, the WFQ classifies packets by their flows, grouping the packets
with the same source IP address, destination address, source port number, destination
port number, protocol number, and IP priority in the IP network into one flow. Each flow is
allocated a queue, and different flows are allocated into different queues as much as
possible. There are at most eight queues. When flows are being sent out of the queue,
WFQ allocates the exit bandwidth to be seized by each flow according to its IP priority.
The smaller the priority value is, the less the bandwidth will be allocated. The larger the
priority value is, the more bandwidth will be allocated, which assures the fairness
between services of the same priority, and the weighted value between services of
different priority. For example, if there are currently eight flows in the interface, with the
priorities 0, 1, 2, 3, 4, 5. 6, and 7, the total bandwidth quantum will be the sum of the
priorities of all the flows + 1 for each flow, that is, 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 = 36. The
bandwidth ratio to be seized by each flow is: (its priority+1)/ (sum of all (flow priorities +
1)), that is, 1/36, 2/36, 3/36, 4/36, 5/36, 6/36, 7/36, and 8/36.
Another example, if there are four flows currently, and the priority of three is 4 and that of
the one is 5, the total bandwidth quantum will be (4 + 1) * 3 + (5 + 1) = 21. In this case,
the bandwidth ratio of the three priority 4 flows is 5/21, and that of priority 5 flow is 6/21.
This evidences the weighted value to the services of different priorities by the WFQ on
fair basis, and the weighted value depends on the IP priority carried in the IP packet
header.
3.19.2.3 Congestion avoidance
Due to limited memory resource, traditionally all arriving packets will be dropped when
the queue length is up to a specified maximum. For TCP packets, dropping large
quantity of packets will result in TCP timeout, which will trigger the -start and congestion
avoidance mechanism of TCP to reduce the rate of sending packets. When the queue
drops packets of several TCP connections at the same time, it will trigger slow-start and
congestion avoidance at these connections at the same time, which is referred to as
TCP global synchronization. In this way these TCP connections will send fewer packets
to the queue, keeping the packets traffic sent to the queue lower than the line forwarding
speed, and reducing line bandwidth utilization. Traffic packets sent to the queue will
keep changing drastically, where traffic on the line keep fluctuating to be either minimal
or fully saturated.
To prevent the above situation from happening, packet drop strategy of Weighted
Random Early Detection (WRED) can be employed, which allow users to set thresholds
for the queues. When the queue length is below the low threshold, WRED drops no
packets. When the length is between the low and high thresholds, WRED begins to drop
packets randomly (the longer the queue is, the higher the probability of being dropped
will be) .when the queue length is longer than the high threshold, it drops all packets.
As WRED drops packets randomly, it prevents several TCP connections from slowing
down their sending speed at the same time, which avoids the global TCP
synchronization phenomenon. When packets of a TCP connection are dropped and TCP
connection begins to slow down its sending speed, the rest of TCP connections still
maintain a high sending speed. In this way there are always TCP connections engaged
in high speed sending, which improves the bandwidth utilization.
If packets are dropped based by simple comparison between instantaneous queue
length and thresholds set by the user (which is the absolute length for setting queue
thresholds), data flows will probably be treated unfairly. Therefore average queue length
is used instead for comparing with the user-configured threshold to decide dropping.
Here the average queue length refers to the result of the queue length being filtered by
36 © 2010 ZTE Corporation. All rights reserved. ZTE Confidential Proprietary