Test Description

1. Parameters:
   • BE packet size
   • EF packet size
2. Stream profiles:
   • EF: 300 Kbps, packet size variable (128 by and 1024 by)
   • BE: 2.0 Mbps, 200 packs/sec, variable packet size: [100, 250, 500, 750, 1000] bytes
3. Test conditions:
   • EF queue-limit = 10 pack (constant)
   • tx-ring-limit: 5 particles
   • EF scheduling algorithm: Priority Queuing
   • Priority Queuing policing rate for EF traffic: 300 Kbps
   • PVC: bandwidth 2 Mbps
4. Router configuration

Results in short:

• given a constant EF packet size and an EF traffic volume which is a small fraction of the link rate, the smaller is the average BE packet size, the less is the EF nodal delay. In fact, for smaller BE packets the time spent by an EF packet into a priority queue decreases.
• given a constant average BE packet size and a constant EF rate, the smaller the EF packet the less is the nodal delay because of two factors: the tranmission time of the EF packet itself decreases and the number of BE packets interleaving EF packets decreases as well (i.e. the number of BE packets at the head of the FIFO tx queue decreases).
• the smaller the EF packet, the higher is the ipdv. This has to be explained!

Comments:

• Figure 1 compares average one-way delay measures for different EF and BE packet sizes.
  Given a fixed EF packet size, average one-way delay increases as a consequence of the increase of the BE packet size. In fact, given an EF service rate which is a small fraction of the line rate (in this case the EF rate / line rate ratio is 0.15), the BE traffic volume is such that EF packets are interleaved by considerable number of BE packets. Since the tx queue is FIFO, if the BE packet size increases, the impact of tx time of BE packets at the head of the tx queue is more relevant and the EF queueing delay increases.

  In addition, given a fixed BE packet size, one-way delay increases with the EF packet size. Since the EF rate is constant, the bigger is the EF packet size, the smaller is the EF packet rate. As a consequence, EF packets are interleaved by a larger number of BE packets.
  Secondly, the increase in one-way delay is due to the EF transmission time, which increases linearly with the EF packet size.

• Figure 2 plots one-way delay over time for different BE packet sizes. In each case the EF stream is constant and equal to ? pack/sec (specify here exact value).
  When the BE packet size is equal to 100 bytes (128 bytes if IP and UDP overhead is included), the tx queue can store up to 5 EF or BE datagrams at a time (one datagram per particle). The minimum one-way delay is equal to 5.18 msec, this value is achieved by a EF packet if an EF packet arrives in the queueing system when the transmission of a BE packet out of the tx queue is just finished. In this case, 4 BE packets are stored in the tx ring, the EF packet is the next datagram selected by the scheduler (the EF queue is a priority queue) and the EF packet is placed into the tx queue. The transmission time of 4 BE packets is 2.54 msec (1 datagram corresponds to 3 ATM cells, the tx time of one BE packet is 0.636 msec). The remaining 2.64 mseconds (5.18 - 2.64) include the segmentation delay, the EF transmission time (approximately 2 msec) and the propagation delay.
  The difference between the maximum and the minimum one-way delay is 1.17 msec. This value approximately corresponds to the tx time of 2 BE datagrams and it could be explained in the worst case scenario: One EF packet is received when a BE packet has just been selected as next candidate for transmission by the scheduler and in the tx queue the transmission of the BE datagram at the head of the queue just started. In this case the queueing delay corresponds to the tranmission time of 5 + 1 BE packets. Is this explanation sound?

  Similarly, for any other BE packet sizes the difference between the maximum and the minimum approximately corresponds to the transmission time of 2 BE packets, as shown in the following table:

  BE payload 	min	max	delta	tx time of 1 BE pack
  size (bytes)	(msec)	(msec)	(msec)	(msec)
  	
  100		5.171	6.320	1.149	0.636	(3 cells)
  250		6.374	8.716	2.342	1.272	(6 cells)
  500		7.708	12.603	4.895	2.112	(11 cells)
  750		10.607	17.312	6.705	3.604 	(17 cells)
  1000		10.306	21.020	10.714	4.664	(22 cells)
  
  

  The minimum value varies with the BE packet size, i.e. with the number of BE packets which are sitting in front of the tx queue when the EF packet arrives, according to the following table:

  BE payload	BE pack rate / 	scheduling		AVG num of pack			
  size 		EF pack rate	order (*)		in tx queue	queueing delay	min delay
  (bytes)							(pack)	(^)	(msec)		(msec)
  
  100		8.56		BBBBBBBBBEBBBBBBBB	4 BE		2.544		5.18
  250		3.07		BBBEBBBEBBBEBBBEBB	3 BE + 1 EF	4.452		6.374
  500		1.62		BBEBBEBEBBEBBEBEBB	2 BE + 1 EF	4.860		7.708
  750		1.14		BBEBEBEBEBEBEBEBBE	2 BE + 1 EF	7.844		10.607
  1000		0.83		BEBEBEBEBEEBEBE		2 BE + 1 EF	9.964		10.306	
  
  

  (*) B = best-effort; E = expedited forwarding
  (^) in the best-case scenario

  The difference between the minimum delay as reported in the table and the queueing delay takes includes the transmission time of the EF packet and the segmentation delay.

• Figure 5 and figure 6 show that ipdv is considerably smaller for large EF packet sizes, whatever BE packet size is deployed.
  For an EF packet size equal to 1024 by (figure 6) the maximum ipdv corresponds to the transmission time of 1 BE packet, as a consequence, ipdv peaks increase with the BE packet size, but the upper bound is limited by the tx time of one MTU best-effort packet.

  On the other hand, for EF packets of 128 bytes (figure 5), the variation in ipdv is considerable. In particular, for small EF packets we have frequent peaks in ipdv, whose height increases with the BE packet size. For example, for BE packets of 1000 bytes, ipdv can be up to 10.63 msec, approximately twice than with EF packets of 1024 bytes and approx. equal to the tx time of 2 BE packets.
  Different EF packet sizes imply different interleaving patterns of BE and EF packets. Can this explain the ipdv experimental results?