Foundry NetIron M2404C and M2404F Metro Access Switches
Configuring QoS (Rev. 03)
Overview
© 2008 Foundry Networks, Inc.
Page 14 of 62
slow-down packet-transmission until all packets arrive at their destinations. Then, responding to
the consequent cessation of packet drops, the TCP hosts will resume their normal transmission
rates.
WRED does not check the protocol (i.e. TCP versus UDP). Since most VoIP installations grant
higher priority to their VoIP traffic, which is typically UDP, WRED is more likely to drop the
lower-priority TCP traffic.
Furthermore, sources that generate a large amount of traffic are more likely to be slowed down
because the relative number of packets that are dropped by WRED per user is directly related to
the user’s portion in the total traffic.
However, if the user has a substantial amount of other UDP traffic, the user should increase its
priority as well, so that the UDP traffic would not be dropped. If the UDP packets are discarded,
not only will the user loses the packets (which must be retransmitted by the application, if they are
retransmitted at all), but also the ability of WRED to prevent the congestion will be impaired, since
UDP does not use the ”slow start“ flow control mechanism that is used by TCP.
WRED is effective only for TCP/IP traffic. Packet sources working with other protocols might not
respond to packet drops or might try to retransmit dropped packets without decreasing their rates.
A schematic illustration of WRED packet dropping is shown in
Figure 8
.
Figure 8: Weighted Random Early Detection
Average Queue Size
The average queue size is based on the previous average and the current size of the queue,
according to the formula:
Average = A * (Current queue size) + B *(Previous average)
Where
A
= 0.5
n
,
B
= 1 – 0.5
n
and
n
, the exponential weight factor, is a configurable integer value
in the range 1 to 15.
As shown in the following chart,
A
, the weight of the current queue size, is 0.5 for
n
=1 and
decreases to approximately
3.05*10^5
as
n
increases to 15.
B
, the weight of the previous average in
this formula is 0.5 for
n
=1 and rises asymptotically to practically
1
as
n
increases to 15.
n
1
2
3
4
5
6 7 8 …
15
A
0.5 0.25 0.125 0.0625 0.03125 0.015625 0.007813 0.003906 … 3.05*10^5