tf-tant: BE test in the wide area

Testing of BE traffic in the wide area
for different BE load values (real packet size distribution)

Goal: to verify the impact of traffic in the network on one-way delay and IPDV in complete absence of any function for traffic differentation. The reason for this is to compare the performance of a best-effort microflow in an ideal scenario (with just 1 stream) with the performance achived by the same stream in presence of background traffic.
This test is to analyze the soundness of the claim that the overprovisioning of bandwidth is the solution to the need of QoS guarantees.

Test Description

Network layout
BE data streams:
- packet size distribution is according to the real distribution computed through the monitoring of traffic from/to the US on a NRN backbone network. The packet size distribution is computed over several billions of samples.
  For each size BE traffic is injected according to such distribution.
- packet size of the BE reference stream: the reference stream is the BE flow generated by the SmartBits to which traffic measurement applies. The packet size is constant and equal to 64 bytes (IP and UDP overhead included).
Router configurations
- no QoS features inabled in the network
Parameters:
- BE load: variable in the range [3, 103]% of the line capacity
Stream profiles:
- multiple BE streams per site
- global number of BE streams is constant
- UDP, each stream is CBR
- reference BE stream: IP packet size = 64 by; rate varies according to overall BE load.
- reference stream performance compared to the equivalent performance achieved by a BE stream without aggregation (only one BE stream)
- Stream profiles:
- no EF traffic
Test conditions:
- PVC: bandwidth 2 Mbps
- queueing algorithm: none, FIFO service policy applied to the transmission queue for EF traffic.
- BE agregation in each site, since

Results in short:

One-way delay is considerably affected by traffic aggregation even when injected load is well below the line capacity (e.g. 58%). We believed that the greater spread in one-way delay is due to packet clustering produced by traffic aggregation.
The maximum one-way delay experienced by the reference stream increases with the overall load.
IPDV distribution is more spread in presence of traffic aggregation in comparison with the equivalent distribution measured without aggregation - 1 stream - however it seems to be independent of the traffic load.

Comments:

Figure 1 and Figure 2 (same as Figure 1, with log scale) plot the one-way frequency distribution by comparing the performance achieved by a reference stream in presence of a variable BE load. One-way delay is expressed in delay units i.e. as a multiple of the minimum delay experienced in the test session (95.37 msec).
One-way delay is strongly influenced by the presence of multiple BE streams, even in presence of a non-congested network. It can be up to 3/2 of the minimum delay. One-way delay gets more spread in a larger range and the average one-way delay increases when the BE load increases, because of the burstiness introduced by packet clustering.
For a BE load of 93% and 103% packet loss is observed on the router located in site 4 (where congestion occurs).
```
BE load (%)	avg one-way delay (msec)


3		99.510
58		101.750
66		104.393
77		104.477
86		114.460
98		115.197
103		128.315
```
Figure 3 and figure 4 compare the IPDV frequency distribution for different BE loads.
IPDV is expressed as a multiple of the transmission time of the reference BE packet (64 bytes, IP overhead included), i.e. 424 microsec.
While the IPDV distribution is strongly influenced by the presence of aggregation (compare the ipdv distribution experienced without aggregation - 1 stream - and with aggregation - all the other curves -), IPDV seems to be independent of the load, as illustrated by the following table.
```
BE load (%)	avg IPDV (msec)


3		0.067
58		3.919
66		4.572
77		4.279
86		4.205
98		4.052
103		3.734
```

Figure 1: BE one-way delay frequency distribution with different BE loads (delay unit = minimum one-way delay)

Figure 2: BE one-way delay frequency distribution with different BE loads (delay unit = minimum one-way delay) - log scale -

Figure 3: BE IPDV frequency distribution with different BE loads (tx unit = tx time of 1 BE reference packet)

Figure 4: BE IPDV frequency distribution with different BE loads (tx unit = tx time of 1 BE reference packet) - log scale -

Last modified: Apr 05, 2000