Dependency of traffic isolation and TCP throughput on queue length
Goal:
To understand the dependency of application performance, in case of
congestion, on the transmit queue size ratio (i.e. of the percentage of
memory allocated to a given queue. Memory has to be shared by all the queues
enabled, for example if we have 3 queues in a 1p2q2t interface, a given
ration can be assigned to each of the low and high priority queues and to
the priority queue. In particular, the ratio allocated to the priority
queue has to be equal to the ratio allocated to the high-priority queue.
Default configuration:
- 1p2q2t:
- low priority: 70%
- high priority: 15%
- strict priority: 15%
- 2q2t:
- low priority: 80%
- high priority: 20%
Equipment:
Network topology:
Router configuration:
Configuration syntax:
wrr queue-limit ratio_queue_1 ratio_queue_2
Traffic profile:
- UDP background traffic:
two solaris workstations are used to inject background UDP
traffic, this to produce congestion on the output interface
of the Cat6500 connecting the switch to router 12000.
Script
- Reference TCP stream: script
Test parameters:
- queue-limit of queue 1 and queue 2
- reference traffic policing at 80 Mbps
- 2 classes: prec 0 (in sub-queue 1,1) and prec 1 (in sub-queue 2,1)
- weight of queue 2,1 (prec 1): 8 (equal to the theoretical weight)
Summary:
- no dependency on queue-limit. Could be due to the large amount of
transmission memory available per port on interfaces like: WS-X6816-GBIC
Comments:
- Figure 1 shows that there is no relationship between TCP performance
expressed in throughput and queue-limit at least for values equal or
larger than 10% of the transmission memory buffer.
Figure 1: application throughput of a TCP stream as a
function of parameter queue-limit.
T.Ferrari and A.Mangiarotti, March 25 2002
