Figure 1 and 2 compare the inter-packet gap frequency distributions of
a single stream at the source and destination side, in two different scenarios:
with policing (policing rate is 30 Mbps) and without policing (the source is
free to get the maximum bandidth available from a GigaEthernet interface). In
both cases just the initial 10000 packets and the corresponding timing information
are logged.
We see that in both cases the distribution shape varies, this implies that
the change in traffic profile is not strictly related to the presence of a policer
but rather by the crossing of the switch fabric.
We can see that in case of no policing the distribution is much more spread
in the lower part of the inter-packet gap range. This is confirmed by the fact
that without policing the transmission rate of a TCP stream is higher.
We also have to be aware of the fact that in the case of no
policing, the high TCP rate
at the source and at the receiving hand can cause holes in packet logging
due to the high transmission rate and possibly by CPU utilization.
The fact that policing has no impact on the single stream
profile is
confirmed by a test we run on a single UDP stream. A single constant bit rate
stream was injected so that the CIR rate of a policer, enabled on the Cat6500
input interface, is not exceeded. The inter-packet gap at the source was
monitored both at the source (before crossing the policer) and the destinatiion
(after policing has been applied to the stream). The corresponding
frequency distributions are plotted in Figure 2.b, as we can see the shape
of the two curves correspond. The good correspondance of the two stream
profiles confirms the correctness of our measurement methodology.
The router configuration used in this test was:
police 25476190 2000000 2000000 conform-action transmit exceed-action drop
The stream configuration at the source and destination was:
sun3 : iperf -c 131.154.100.4 -u -b 25m -t 3600 -p 50006 -i 10 &
// 25.476190 Mbps is the overall rate generated - with IP and UDP
// overhead included
sun4 : iperf -s -p 50006 -u -i 10
Figure 1: inter-packet gap frequency distributions at source and destination
with policing
^M
Figure 2: inter-packet gap frequency distributions at source and destination
without policing
Figure 2.b: inter-packet gap frequency distribution at source and
destination of a CBR UDP stream with policing
We have analyzed the profile of the TCP stream with policing in greater
detail by plotting and comparing the inter-packet gap at transmission and
reception and by analyzing the stream profile over time through the tool
tcptrace.
Figure 3 shows how the profile varies over time. The red curve corresponds
to the traffic profile at the source, while the blu one is the traffic profile
as received by the destination. We can observe the following characteristics:
Figure 3: comparison of inter-packet gap profile at source and destination
over time
- at the source: inter-packet gaps show periodic peaks which corresponds to
the time interval needed to recover from packet loss. Multiple packets are lost
in a burst, this is probably due to the normal burst size configured, which
restricts the burstiness injected by the source.
The presence of periodic burstly loss is confirmed by Figure 4.
Figure 4: packets transmitted over time, periodic loss and retransmission
occur.
- inter-packet gap at the source tends to increase between two consecutive
peaks. The low inter-packet gap may correspond to the slow start phase which
follows the expiration of a timeout, during which the instantaneous rate
is larger than the average rate sent during congestion avoidance. The slow
start phase is illustrated in Figure 5.
Figure 5: slow start phase after expiration of a timeout.
- Figure 6 shows the effect of normal burst size saturation due to the fact
that traffic is injected at a rate higher than the rate at which acknowledgements
are transmitted back to source.
Figure 7
