tf-tant: EF burstiness as a function of the EF packet size

EF Burstiness as a Function of the EF packet size

Goal: measurement of the relationship between EF packet clustering and the size of the EF packet size in a network in which the departure rate is smaller than the maximum instantaneous EF arrival rate.

Test Description

Network layout
BE data streams: single BE stream injected by INFN to congest the output interface:
- BE payload size: 1000 by;
- BE pack rate: 200 pack/sec (2 Mbps); CBR
- BE protocol: UDP
Example of router configuration
Parameters:
- EF IP payload size: [40,240] bytes, according to the typical data frame sizes for voice over IP.
  For a survey on typical IP packet sizes used by codecs for Voice over IP see here (Word document)
EF stream profiles:
- EF load: constant, equal to 640 Kbps, i.e. 120 kbps per site
- EF streams: 10 per site, 40 in total
- 1 EF stream generated by the SmartBits
- EF protocol: UDP, CBR
- EF rate: same for each stream, same EF packet size for each stream
- examples of scripts per test site:
Test conditions:
- PVC: bandwidth 2 Mbps
- queueing algorithm: priority queuing applied to EF
- EF agregation and congestion in each transit site
- 10000 samples per test
Test methodology
The number of EF sreams varies in the range [1, 100] and each site injects the same amount of traffic.
For each test in each router on the data path we measure the amount of EF packets tail dropped by the priority queue (which serves EF packets) and we progressively increase the priority queue size Q until no packet loss caused by tail drop is observed during the whole test. Given the queue size Q at the time when such condition applies, we assume that the maximum burst size (in packets) is Q -1. The burst size in bytes can be derived since the EF packet size is constant and known.
Measurement is applied to a single stream, which is called the reference stream.

Results in short:

burstiness: increases with the EF packet size (if expressed in bytes). It decreases (as exptected) if it expressed in packets, since in this test the EF rate is constant (and a smaller EF packet implies a larger EF packet rate).
one-way delay: increases with the packet size, i.e. with the packet rate. The increase is not due to the (minor) difference in burst size experienced by different size packets but to the tx queue, whose composition - in terms of number of BE and EF packets - varies with the EF packet rate.
IPDV distribution gets more spread with small EF packet sizes (?)

Comments:

As Figure 1 shows, the burst length (expressed in bytes) increases with the EF packet size (EF load is always constant).
Can a formula which expresses the probability of packet clustering as a function of the EF packet size and rate be derived?
However, the burst is longer with small EF packets if the length is expressed in packets (as expected). As a consequence, with small EF packets the EF queue has to be tuned in order to avoid EF packet drop. Since queue memory is allocated in packets, if the EF packet is small, the memory gets fragmented and the EF queue length has to be large enough to avoid packet drop.
Figure 2 (a) and (b) (linear and logarithmic scale respectively) plot the one-way delay distribution for different EF packet sizes. Delay is expressed in delay units, where the delay unit is the minimum one-way delay observed and in this test corresponds to 113.89 msec.
Both graphs indicate that one-way delay increases with the EF packet size and values get more concentrated in an interval around the average (see the table below) when the EF packet size increases.
Two main reasons explain this behaviour:
- the tranmission time increases with the EF packet size, for example the transmission times in the two extreme cases are the following (they are computed by taking into account the ATM overhead):
  - 40 bytes: 0.424 msec
  - 240 bytes: 1.272 msec
- the contribution of the tranmission queue to one-way delay changes with the EF rate (in pack/sec):
  - BE rate: 214 pack/sec
  - EF rate (40 bytes): 720 pack/sec, 3.3 EF packs every BE pack
  - EF rate (240 bytes): 240 pack/sec, 1.1 EF packs every BE pack
  Let's assume that no burstiness is observed.
  In this case, the component of the tx queue is the following:
  - EF 40 bytes: BEBEB, tx time = 1.2 * 2 + 4.6 = 7.0 msec
  - EF 240 bytes: BEEEB, tx time = 0.424 * 3 = 1.2 msec
This difference in delay contribution is introduced by each tx queue on the data path, i.e. by each router which experienced congestion.
```
EF packet size (bytes)	Avg one-way delay (msec)

40			132.629
80			134.834
120			141.325
240			144.715
```
Figure 3 and 4 plot the IPDV distribution for different EF packet sizes (in linear and logarithmic scale respectively).
According to both of them the IPDV distribution gets more spread for smaller packet sizes. This could be justified by the burst length (as explained in Figure 4 in the test with variable EF rate): with longer burst sizes IPDV values are more concentrated around the average.
Discussion needed on this result
```
EF packet size (bytes)	Avg one-way delay (msec)

40			8188
80			6608
120			4879
240			5416
```

Figure 1: EF burstiness for different numbers of EF streams

Figure 2 (a): one-way delay frequency distribution for different EF IP payload sizes

Figure 2 (b): one-way delay frequency distribution for different EF IP payload sizes (logarithmic scale)

Figure 3: IPDV frequency distribution for different numbers of EF streams

Figure 4: IPDV frequency distribution for different numbers of EF streams (logarithmic scale)

Last modified: Apr 10, 2000