tf-tant: EF in the wide area

Testing of EF in the wide area
in presence of stream aggregation and congestion

Goal: to verify the relatinship between one-way delay and ipdv (for different EF packet sizes) and:

aggregation of 4 EF UDP streams at different rates and with different packet sizes,
congestion in 4 consecutive points on the data path,
priority queuing applied in a chain of four routers,

Test Description

Network layout
EF and BE data streams:
- sources and destinations
- command lines
Router configurations
Parameters:
- traffic aggregation:
  - Scenario 1: 1 EF stream without BE
  - Scenario 2: 4 EF streams withoutBE
  - Scenario 3: 4 EF streams with BE
- packet size of 1 EF stream (generated by SmartBits)
Stream profiles:
- measured EF stream: constant bit rate equal to 300 Kbps, packet size variable
- additional EF traffic (if present):
  - total EF rate equal to 1.2 Mbps (ATM overhead included)
  - EF streams:
```
EF stream	EF rate		EF packet size (bytes)* 

SmartBits	300 Kbps		variable
from CERN	400 Kbps		1024
from GRnet	300 Kbps		512
from Utrecht	200 Kbps		200
		
```
    (*) size does not include IP and UDP overhead.
  - BE stream: constant bit rate equal to 2.0 Mbps (217 packs/sec, 1000 bytes/pack)
Test conditions:
- EF queue-limit = 10 pack (constant)
- tx-ring-limit: 5 particles (constant)
- PVC: bandwidth 2 Mbps
- queueing algorithm for EF traffic: PQ, policing rate: 1.2 M
- policing rate at ingress interfaces equal to input EF rate, exceed action: drop, burst tolerance: 8000 bytes

Results in short:

end-to-end one-way delay is proportional to the number of congestion points on the data path from source to destination.
the greatest contribution to one-way delay derives from the queuing delay introduced by the transmission queue.
aggregation causes an increase in both one-way delay and ipdv also without congestion! This is due to packet clustering on the data path.
the impact of aggregation on ipdv is proportional to the probability that packets get clustered, i.e. it is inversely proportional to the packet size of a stream (when testing with different packet sizes and constant output rate).
for any packet size the one-way delay frequency distribution has the shape of a normal distribution.
The ipdv frequency distribution shows that for larger packets ipdv is more spread but the most frequent ipdv value is lower than with small-size packets.

Comments:

Figure 1 compares the average one-way delay measured in the ideal scenario, i.e. with just one EF stream (no aggregation) and no best-effort traffic to the end-to-end delay measured:
1. with aggregation and without congestion;
2. with aggregation and with congestion;
Aggregation introduces a minor delay increase which is maximum for small packet sizes (6.78 msec) and gradually decreases with the packet size: Given a constant rate (300 Kbps) for small packet sizes the number of packet issued per time interval increases, as a consequence in a given aggregation point the probability that in one time interval one or more EF packets are enqueued at the same time of similar EF packets sent out by other EF sources increases.
However, the greatest contribution to the increase in end-to-end delay derives from the nodal delay accumulated in each diffserv node on the data path. This delay is independent of the EF packet size and is a linear function of the number of routers in the chain. For example, for a frame size equal to 64 bytes the delta between the blue and pink curve is 37.21 msec. As test 7 shows, the minimum nodal delay added by the diffserv router when the tx queue lenght is 5 packets, is equal to the transmission time of 2 BE packets (1000 bytes, 4.66 msec per packet), i.e. 9.32 msec. By multiplying this delay times the number of routers on the path we get 37.28 msec, which is almost equivalent to the one-way delay increase measured.
We can conclude that given the nodal delay D_i introduced by the i^th diffserv router, the overall one-way delay D introduced by a chain of n can be simply computed as:
D = D₁ + D ₂ + ... + D _n
Figure 2 plots the average ipdv measured in the three different test scenarios.
Without aggregation and congestion ipdv is almost negligible, as expected. However, even without background traffic (i.e. without congestion) the addition of 3 EF streams introduces a considerable ipdv which depends on the EF packet size.
Since in this case there is no congestion, the transmission queue should always be empty and ipdv is probably introduced just by the priority queue. ipdv can only be explained by the presence of multiple packets in the priority queue at a time, i.e. by packet bursts. We think that bursts could be introduced when packets from different EF streams and different input interfaces are aggregated into the same class and sent to the same output queue. Bursts are sent out of the priority queue at line rate. Bursts probably tend to get longer through the data path when multiple additional aggregation points are met. In fact, the longer is an imput burst, the longer is the transmission time, so the higher is the probability that packets from a different input stream find a burst at the head of the priority queue.
ipdv tends to decrease for larger packet sizes (only the packet size of the stream measured varies, while the EF background traffic is contant both in rate and packet size). Since the EF reference stream is sent at constant bit rate, for larger packet sizes the number of packets issued per time interval decreases. As a consenquence, the potential number of bursts decreases with the packet size.
The EF behavior for different EF packet sizes is also illustrated in figure 3 and figure 4.
The one-way delay frequency distribution is the approximation of a normal distribution for any packet size (e.g. 128 bytes - figure 5 and 1518 bytes - figure 6).
Figure 7 and figure 8 compare the ipdv frequency distribution for two packet sizes: 256 and 1518 bytes respectively. The ipdv range is divided into equal lenght intervals. For 256 bytes ipdv values are more concentrated around the average (range: [4.452, 5.934] msec) but the most frequent ipdv value is higher than for 1518 bytes frames. In the other case, ipdv is more spread, but the most frequent value is close to the ideal ipdv value, which is in the range [0.020, 1.103] msec.
So large packet size experience better performance.
The ideal one-way delay and ipdv frequency distributions (the one measured for a single EF stream without both aggregation and congestion) is illustrated in figure 9 and figure 10.

Figure 1: average one-way delay of a single EF stream, of the same stream in presence of 3 additional EF streams with and without congestion (produced by a BE stream).

Figure 2: average ipdv for different EF packet sizes in three scenarios: with just one EF stream (baseline test), with EF aggregation (without congestion) and with both aggregation and congestion (4 EF streams plus a BE stream).

Figure 3: one-way delay over time for a single EF stream with constant rate equal to 300 Kbps and constant frame size equal to 256 bytes. The same measurement is repeated in three different scenarios.

Figure 4: one-way delay over time for a single EF stream with constant rate equal to 300 Kbps and constant frame size equal to 1024 bytes. The same measurement is repeated in three different scenarios.

Figure 5: one-way delay frequency distribution for EF frames of 128 bytes within a single EF stream.

Figure 6: one-way delay frequency distribution for EF frames of 1518 bytes within a single EF stream.

Figure 7: ipdv frequency distribution of EF packets within a single EF stream with constant frame size equal to 256 bytes. In each interval the number corresponds to the lower bound of the range.

Figure 8: ipdv frequency distribution of EF packets within a single EF stream with constant frame size equal to 1518 bytes. In each interval the number corresponds to the lower bound of the range.

Additional pictures:

Figure 9: ideal one-way delay frequency distribution of EF packets belonging to the same stream without either aggregation or congestion (EF frame size = 256 bytes).

Figure 10: ideal ipdv delay frequency distribution of EF packets belonging to the same stream without either aggregation or congestion (EF frame size = 256 bytes).

Figure 11: 50, 90 and 95 percentile of one-way delay with aggregation and congestion.

Last modified: Feb 01, 2000