tf-tant: EF burstiness as a function of the EF rate

EF Burstiness as a Function of the EF Rate

Goal: measurement of the relationship between EF packet clustering and EF load in a network in which the departure rate is smaller than the maximum instantaneous EF arrival rate.
In some network scenarios, it can happen that the departure rate cannot be greater or equal than the maximum instantaneous arrival rate as stated by RFC 2598. If such requirement cannot be met, then the arrival rate can produce EF packet clustering and instantaneous non-empty EF queues, i.e. the formation of EF bursts which propagate on the data path to the destination.

The instantaneous EF arrival rate can be larger than the max departure rate when multiple high-speed LAN interfaces inject multiple EF streams which are sinked to the same ouput interface, as in this test scenario (see Figure 1).
In this test we measure EF burstiness produced by packet clustering, as a function of the EF load.

The test scenario is illustrated in Figure 1.

Figure 1

Max Arrival Rate Estimation
Generally speaking, given n input EF BA, each injected by a different input interface i_j with line rate l_j, by assuming that the MTU size is the same for each interface, then the maximum arrival rate A_r is:

A_r = (n * MTU_size * 8) / min_tx_time (MTU)

min_tx_rate(MTU) = (MTU_size * 8) / max(l_i)

So we can conclude that: A_r = n * max(l_i)

Aggregation Degree of BA

A = 1 - max(r_i) / r_BA where:

r_i: rate of i^th stream of a given BA
r_BA: overall traffic load for behavior aggregate BA

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 pack7sec (2 Mbps); CBR
- BE protocol: UDP
Example of router configuration
Parameters:
- EF load: variable in the range [200, 1000] Kbps, i.e. [10, 50]% of line rate
Stream profiles:
- 10 EF streams per site (40 in total), 1 EF stream generated by the SmartBits
- EF protocol: UDP, CBR
- EF rate: same for each stream, constant EF IP packet size: 68 bytes
- 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
- 5000 samples per test
Test methodology
We define EF laod the overall traffic volume produced by EF streams. The EF load varies in the range [200, 1000] Kbps 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:

EF load has a great impact on EF burstiness, which varies linearly with the EF load produced.
the queuing delay introduced by EF queues in presence of EF burstiness is such that the delay contribution is not frequent enough to produce a measurable effect - this test has to be double-checked.
the effect of EF burstiness on IPDV is such that for smaller EF load values IPDV increases.

Comments:

Figure 2 plots the EF burstiness measured for different EF load values (in the range [10, 50]% of the line rate).
We see that burstiness is approximately a linear function of the rate. Burstiness arises when packets from different interfaces and belonging to different EF streams are sinked to the same output interface: In this test scenario in a given test site streams are generated by nodes upstream - they arrive on a 2 Mbps input interface - and from the local LAN interface (10Mb/100Mb or OC-3c depending on the test site).
EF instantaneous burstiness can produce packet loss, in particular when the EF packet size is small, since queues are allocated in packets and for small packet sizes the memory allocation scheme is highly inefficient (memory is fragmented and the queue is more easily subject to overflow).
In order to avoid packet loss caused by EF burstiness, the EF queue needs size needs to be properly dimensioned (EF queue size > burst size).
In Figure 2 (b) the Ef burstiness measured with PQ is compared to the equivalent burstiness with WFQ.
The test shows that with WFQ the burst length is slightly smaller than with PQ, and the difference between the two is the same independently of the EF rate.
A better WFQ performance in terms of burstiness is expected, since with WFQ the transmission of EF packets can be interrupted by the scheduler if the BE packet at the head of the BE queue has a forwarding time which is less then the time of the first EF packet in the EF queue. The WFQ algorithm helps at preventing the formation of very long bursts.
Figure 3 plots the one-way delay frequency distribution, where one-way is expressed in delay units. We define delay unit as the minimum one-way delay value experienced by the reference stream, where the minimum is computed over all the tests performed.
In this test the one-way delay unit is equal to 108.14 msec.
The increasing EF queue length which is configured in order to avoid packet loss has a small effect on one-way delay: One-way delay slightly decreases as Figure 3 shows, this is clear from the table below which indicates the average one-way for each EF load. One-way delay slightly decreases, a result which is against our intuition (this test needs to be repeated).
```
EF load		Avg one-way delay (msec)

10%		131.768
20%		130.879
30%		128.909
40%		128.451
50%		128.470
```
In the worst case scenario (burst size=35 pack) the queuing delay introduced by the priority queue is up to 14.84 msec. However, if the presence of long EF bursts is occasional, the effect is not very visible from distributions curves, still it could have a negative impact on the end-to-end application performance.
Figure 4 plots the IPDV frequency distribution for different EF loads (the tx unit is equal to the tx time of one EF packet, i.e. 0.424 msec). The curves show that in presence of increasing EF load values, i.e. of increasing EF burstiness and consequently of longer EF queues, the IPDV distribution gets more spread around the average and the maximum IPDV observed increases as well.
The following table, indicates the sample mean for different EF load values.
```
EF load		Avg one-way delay (microsec)

10%		6665
20%		9712
30%		8772
40%		5672
50%		6858
```
A possible interpretation of this is that for larger rates, packets tend to cluster into longer bursts. Best-effort packets are transmitted only when the priority queue is empty. This implies that the longer is the burst, the greater is the number of packets which are transmitted at the same run by the priority queue without being interleaved by BE packets. This also implies that for a greater number of EF packets the delay experienced in a PQ does not considerably differ from the corresponding delay of the previous and subsequent packet in the burst, i.e. IPDV is more constant.

Figure 2: EF burstiness for different EF load values

Figure 2 (b): comparison of EF burstiness with WFQ and PQ

Figure 3: one-way delay frequency distribution for different EF load values
(i.e. of different burstiness degress, that is of different EF queue sizes)

Figure 4: IPDV frequency distribution for different EF load values
(i.e. of different burstiness degress, that is of different EF queue sizes)

Last modified: Apr 10, 2000