DESCRIPTION OF TEST A

• 1 source: INFN
• 4 destinations: Switch (800 Kbps), Rediris (400 Kbps),
  DANTE (200 Kbps), Uni Stuttgart (100 Kbps)
• 4 TCP streams
• per-vc CBWFQ configured in the ingress router, output interface:
  PA-A3. 4 classes one per destination, classification based
  on source IP address and destination IP network address

REMARKS

• CBWFQ takes effect only under congestion, if a single TCP stream is
  generated (no resource contention), the stream can get the total
  amount of bandwidth (as expected)
• when a single TCP stream (socket size 250000 bytes) is generated per class
  (4 steams in total), no real congestion is achieved and each streams
  gets an amount of bandwidth which is not proportional to the weight
  of the correponding class, this amount of bandwidth varies and can be
  higher or lower than the corresponding bandwidth.

DESCRIPTION OF TEST B

• 1 source: INFN
• 1 destination: Switch (800 Kbps)
• same router configuration as above
• 1 UDP stream (best-effort) of 150 pack/sec, 1000 by per pack
• incresing number of TCP connections to Switch (from 1 to 15)

REMARKS

• the amount of bandwidth used as weight in the CBWFQ configuration is
  is much higher than the throughput of a single TCP connection. When
  the number of TCP connections increases the aggregate throughput increases
  up to 796.8 Kbps (this is the threshold, throughput is at the application
  layer (does not include overhead)
• the UDP stream is only used to generate congestion, pack loss increases
  when the number of TCP connections increses
  CBWFQ drops packets only at the start of the TCP streams (they are
  concurrent and start at the same time). After this phase TCP adapts and
  packet dropping do not occur anymore

Detailed configuration of the router
at INFN, no QoS features enabled on other routers.

• 1 source: INFN
• 4 destinations: Switch, Rediris, DANTE, Uni Stuttgart
• 4 UDP streams (mgen), 1 to each destination at a constant rate:
  150 pack/sec, 1000 by per pack, UDP traffic is shaped

REMARKS

• without CBWFQ each streams gest alomost an equal share of bandwidth
  (around 378 Kbps each)
• with CBWFQ each receiver receives UDP datagrams at a rate which is
  *exactly* equivalent to the bandwidth associated to the class (as expected,
  good result!), this also imply that pack loss is inversely proportional
  to the wight of the class
• the addition of background TCP traffic (10 streams) does not modify the
  UDP performance of each stream. The aggregate TCP throughput oscillates
  in the range [300, 420] Kbps
• The sum of the aggregate UDP and aggregate TCP throughput gives 1.9 Mbps
  approximately (good result, since overhead is not included)

July 8, 1999