Test Description

1. Parameters:
   • BE packet size distribution:
     1. deterministic (constant BE packet size):
        • Example of script for 450 by packets
     2. geometric (for different average values): example of scripts
        • 128 by
        • 512 by
        • 1024 by
     3. real (according to packet size monitoring applied to traffic on an intercontinental connection (in both ways):
        • script
2. Stream profiles:
   • EF: 300 Kbps (constant rate), frame size constant: 128 by
   • BE: overall rate > 2.0 Mbps, 20 CBR streams
   • BE packet size: constant for a given stream. Packet size according to 3 different traffic distributions.
   • BE: UDP
3. Test conditions:
   • EF queue-limit = 10 pack (constant)
   • tx-ring-limit: 5 particles
   • Priority Queuing policing rate = 300 Kbps
   • PVC: bandwidth 2 Mbps
4. Network layout

Results in short:

• Best-effort traffic pattern has an influence on EF one-way delay, because of the presence of the tx queue, which is partially allocated to BE traffic in presence of congestion.
• EF one-way delay is distributed over a smaller range in presence of smaller BE packet sizes, in fact in this case the queuing delay introduced by both the BE queue and the tx queue is smaller.
• With a real BE packet size distribution one-way delay distribution is more spread, given the presence of BE packets of very different size in the BE queue: 64 bytes or MTU-size packets.

Comments:

• Figure 2 plots the EF one-way sample mean for increasing BE packet sizes. In each test, for each BE stream the packet size is constant and unique for each flow.
  This test shows the effect of the transmission queue on one-way delay: the curve does not increases linearly because of the different contribution of the tx queue to queuing delay: The queuing delay depends on how much of a particle (512 bytes) is allocated to a packet. For BE packets whose length is an integer multiple of particle, we see a maximum (around 500 and 1000 bytes), the one-way delay decreases since for larger packets size the particle allocation is less efficient (part of the tx memory is not allocated and the corresponding time needed to empty it is lower).
  The effect of the tx queue is also shown by Figure 3, which plots one-way delay over time for two given BE packet sizes: the two curves as similarly shaped but the minimum is different.
• Figure 4 shows that generally speaking, even in presence of more complex packet size distributions (geometric in this case), not only the average but also the one-way delay frequency distribution shape changes. for larger BE packet sizes one-way delay is distributed over a larger range.
  EF performance is better in presence of smaller BE packet sizes, in fact in this case the queuing delay introduced by both the BE queue and the tx queue is smaller.
• Figure 5 shows that in presence of a real BE packet size distribution (in which small and MTU size packets are the most frequent one) one-way delay varies in a larger range than with a geometric distribution (due to the great difference in BE packet sizes).
• Given the frequency distribution (Figure 5), the equation of the inverse probability function of the queuing delay introduced by an EF queue (plus a FIFO tx queue) can be computed (Figure 6), so that the probability that delay is larger than a given values can be determined. This can be useful at the application layer for playback buffer dimensioning.