Luis Sequeira, Julian Fernandez-Navajas, Jose Saldana, Luis Casadesus, Jose Ruiz-Mas, "Empirically Characterizing the Buffer Behaviour of Real Devices," Proc. International Symposium on Performance Evaluation of Computer and Telecommunication Systems SPECTS 2012, July 8-11, 2012, Genoa, Italy. ISBN: 978-1-4673-2235-5.

1. Empirically Characterizing the
Buffer Behaviour of Real
Devices
Luis Sequeira
Julián Fernández-Navajas
Jose Saldana
Luis Casadesus
Communication Technologies Group (GTC)
Aragón Institute of Engineering Research (I3A)
University of Zaragoza, Spain

2. Index
I. Introduction and Related Works
II. Test Methodology
• Test Procedure
• Methodology
III. Experimental Results
• Wired Scenario
• Wireless Scenario
IV. Conclusions

3. Index
I. Introduction and Related Works
II. Test Methodology
• Test Procedure
• Methodology
III. Experimental Results
• Wired Scenario
• Wireless Scenario
IV. Conclusions

4. Introduction and Related Works
• Packet size varies for different Internet
services and applications:
• VoIP, web, FTP, video streaming, online games, etc.
• Traffic behaviour has to be taken into account
for the design of network devices:
• Constant bit rate traffic
• Bursty traffic
• Different number of frames into each burst

5. Introduction and Related Works
• In access network capabilities are lower than
the ones available in the backbone; this
implies that some network points become
critical bottlenecks
• Bottlenecks may also appear at critical points
of high-performance networks.

6. Introduction and Related Works
• Mid and low-end routers normally do not implement
advanced traffic management techniques.
• They always use buffers as a traffic regulation
mechanism.
• Buffer size becomes an important design parameter
• Buffer can be measured in different ways:
• Maximum number of packets it can store
• Amount of bytes
• Queuing time limit (ms)

7. Introduction and Related Works
• Most Internet routers use FIFO, but there exist
other techniques to manage drop-tail buffers,
e.g. Random Early Detection (RED).
• These techniques, in conjunction with buffer
size, mainly define the buffer behaviour and
how traffic is affected by it.

8. Introduction and Related Works
• Relationship between router buffer size and
link utilization:
Excessive amount of memory
latency increase
Small amount of memory
packet loss increase
• Characterization of buffer behaviour is
interesting when trying to improve link
utilization.

9. Index
I. Introduction and Related Works
II. Test Methodology
• Test Procedure
• Methodology
III. Experimental Results
• Wired Scenario
• Wireless Scenario
IV. Conclusions

10. Test Procedure
It may be a device
or a network

11. Test Procedure
UDP Packets of different size are used
to determine if buffer size is in bytes or
in packets
to produce a buffer overflow in
the SUT
All packets have a
unique ID

12. Test Procedure
Traffic is captured here

13. Test Procedure
Input rate
Buffer of the
“System Under Test”
Input rate > output rate
Output rate

14. Test Procedure
Input rate
Buffer of the
“System Under Test”
Input rate > output rate
Buffer fills
Output rate

15. Test Procedure
Filling the buffer
Input rate
i
2
1
Output rate

16. Test Procedure
Time to fill the buffer
First loss
Input rate
i
2
1
Packet i leaves the
buffer
Error
Time to fill and to
emtpy the buffer
Output rate

17. Methodology
• Physical Access:
Counting the
number of
packets in the
queue in the
moment that a
packet arrives to
the buffer.
Traffic may be captured here
packet delay
packet loss
interarrival packet time
input and output buffer rate
filling buffer rate
captures
script
BUFFER SIZE

18. Methodology
• Remote Access:
If the delay of a
packet in the
buffer can be
determined, then
the variations of
this delay can
give us useful
information for
estimating buffer
size.
Traffic may be captured here
packet delay
packet loss
interarrival packet time
input and output buffer rate
filling buffer rate
captures
script
BUFFER SIZE

19. Methodology
Input rate
Method 1: Physical Access
Time when the packet
i arrived to the buffer
i
2
• Buffer size is determined for each packet as follows: for
all packets in out-capture, a shell script looks for the
incoming time in in-capture and counts in out-capture
the number of packets between incoming time and the
time stamp registered in out- capture, finally the buffer
size is estimated as the average of all these values.
1
Output rate

20. Methodology
Input rate
Method 1: Physical Access
Buffer gets empty
• Buffer size is determined for each packet as follows: for
all packets in out-capture, a shell script looks for the
incoming time in in-capture and counts in out-capture
the number of packets between incoming time and the
time stamp registered in out- capture, finally the buffer
size is estimated as the average of all these values.
i
Output rate

21. Methodology
Input rate
Method 1: Physical Access
Time when the
packet i leaves
the buffer
• Buffer size is determined for each packet as follows: for
all packets in out-capture, a shell script looks for the
incoming time in in-capture and counts in out-capture
the number of packets between incoming time and the
time stamp registered in out- capture, finally the buffer
size is estimated as the average of all these values.
i
Output rate

22. Methodology
Input rate
Method 1: Physical Access
Queue
Count packets
Queue
• Buffer size is determined for each packet as follows: for
all packets in out-capture, a shell script looks for the
incoming time in in-capture and counts in out-capture
the number of packets between incoming time and the
time stamp registered in out- capture, finally the buffer
size is estimated as the average of all these values.
i
Output rate

23. Methodology
Method 2: Remote Access
It will completly
fill the buffer
Input rate
i
2
1
Dropped
packets
Output rate

24. Methodology
Method 2: Remote Access
Input rate
It will completly
fill the buffer
i
2
𝑅 𝑓𝑖𝑙𝑙
1
Output rate
Time to fill the buffer

25. Methodology
Method 2: Remote Access
Input rate
Time to fill and
empty the buffer
𝑅 𝑜𝑢𝑡
𝑅 𝑓𝑖𝑙𝑙
i
Output rate
It appears in the
receiver

26. Methodology
Estimation of the size:
Time to fill
the buffer
Time to fill and to
empty the buffer

27. Index
I. Introduction and Related Works
II. Test Methodology
• Test Procedure
• Methodology
III. Experimental Results
• Wired Scenario
• Wireless Scenario
IV. Conclusions

28. Wired Scenario
• In this case we want to test
the accuracy of our method
when there is no possibility
of physical access to the
SUT.
• We obtained the buffer size
using method 1 with
physical access. Next, the
estimations obtained using
method 2 are compared
with previous results, and
the relative error is
obtained.

29. Wired Scenario
A Particular buffer behaviour:
Linksys WAP54G
Groups of dropped packets
3COM 4500

30. Wired Scenario
A Particular buffer behaviour:
• It has been observed in the wired
and wireless tested:
• when the buffer is completely full, no
more packets are accepted, it will be
called upper limit.
• The buffer does not accept new
packets until a certain amount of
memory is available, it will be called
lower limit.
• In this moment the filling process
begins again.
• Although this behaviour has some
similarities with Random Early
Detection (RED) but it is not the
same:
• There is not probability for dropping
an incoming packet.

31. Wired Scenario
• Three different amounts of
bandwidth have been used in
order to flood the buffer.
• We see that the accuracy of the
buffer size estimation using the
method 2 is high. In addition, the
error diminishes as input rate
grows.
• The results are less accurate but
they are still acceptable.
• The results using smaller packets
are less accurate so we have not
presented them.
(packet size = 1500 bytes)

32. Wireless Scenario
• Variations of the output
rate in wireless network
generate error growth.
• While filling rate is relatively
constant, the emptying rate
shows variations for the
highest bandwidths.
The WiFi access point
switches from higher to
lower speeds depending on
the status of the radio
channel

33. Wireless Scenario
• We have compared the two
methods when there is
physical access to the SUT.
• The presented results are
the ones obtained using
packets of 1500 bytes, since
they are the most accurate.
• Method 1 is the most
accurate estimation so it
has been used to compare
with method 2.

34. Index
I. Introduction and Related Works
II. Test Methodology
• Test Procedure
• Methodology
III. Experimental Results
• Wired Scenario
• Wireless Scenario
IV. Conclusions

35. Conclusions
• Two methods have been presented in order to analyze the
technical and functional characteristics of commercial buffers of
different devices, or even networks.
• This characterization is important, taking into account that the
buffer may modify traffic characteristics, and may also drop
packets.
• The methodology can be used if there is physical access to the
“System Under Test”, but it is also useful, with certain limitations,
for measuring a remote system.
• Tests using commercial devices have been deployed in two
different scenarios, using wired and wireless networks.

36. Conclusions
• A particular buffer behaviour has been observed for a device: once
the buffer is full, it does not accept new packets until a certain
space is again available.
• The results show that accurate results of the buffer size can be
obtained when there is physical access to the “System Under Test”.
• In case of having no direct access to the system, an acceptable
estimation can also be obtained if the input rate is more than three
times the output rate. In this case, big packets have to be used for
the tests.
• As future work the method has to be improved in order to
minimize the error, especially when measuring wireless devices.