Goal: to verify the actual burstiness degree produced by a variable number of TCP concurrent streams in different test scenarios: with and without ECN and for different drop probabilities.
Equipment:

• chain of intermediate ecn-capable routers (Cisco 7500 and C7200), IOS: 12.1(4)E (experimental version) - only first router is ecn-capable in this test
• pool of Linux hosts, v. 2.4.9, SACK-capable (no Solaris hosts in use)

Test description:

1. Tolopology see Test A.
2. Parameters
   • min threshold constant = 100 pack
   • max threshold constant = 300 pack
   • drop probability: [1/50, 1/5, 1/2]
   • TCP stacks: ECN and SACK-capable and non-ECN and SACK-capable (Linux hosts only)
   • test duration: 3 min
   • Variable number of TCP streams (streaming mode): [4,8,32,64]
   • minimum link capacity available: 50 Mbps.
3. Traffic profile
   • TCP streams generated by iperf;
   • no background streams, in the sense that all are ECN or non-ECN capable;
4. Test methodology
   SNMP has been used to query the ECN capable router every 2 seconds. Two are the variables tracked: SnmpCurrQ, SnmpMeanQ, both from the Cisco-class-based-qos-mib.
   When monitoring the two parameters we verified that only every 8 or 10 seconds the two parameters are updated, this means that 8-10 sec is the minimum monitoring granularity achieved.
   This is the script used to query the router.

List of tests:

• non-ECN, variable number of streams (drop probability=1/50)
• non-ECN and ECN (64 streams, drop probability=1/50)
• variable drop probability (ECN, 64 streams)

Summary:

• both with ECN and non-ECN streams we can see that an increase in number of parallel streams greatly influences both the instantaneous and the average RED length, as also shown in literature;
• if ECN is enabled the overall burstiness of the aggregate stream is reduced, this is a proof of the correct behavior of ECN and of its positive role in congestion control; in fact, the effect of ECN on burstiness control increases with the drop probability;
• the averaging behavior built in WRED is correclty shown by the comparison of the mean and instantaneous queue lengths: the mean queue length variation is much smoother than in the latter case.

Comments:

• Comparison for a variable number of TCP streams (non-ECN capable, drop probability=1/50):
  In this test we simply perform an analysis of the aggregate TCP stream without enabling ECN. The purpose of this test was to verify if the number of active TCP streams has an impact on the overall burstiness resulting from the multiplexing of a number of indipendent TCP streams.

  Variation in aggregate instantaneous and average burstiness is shown in the following table. The corresponding graphs are in Figure 1, 2, 3 and 4.

  Num of streams	Max instant. queue	Num. of intervals of 10 sec with non-zero inst
  		length (packets)	and average queue length over 180 sec
  -------------------------------------------------------------------------------------
  4		100			6  intervals / 18
  8		100			10 intervals / 18
  32		110			13 intervals / 18
  64		160			always non-zero	
  

  Figure 1: instantaneous and average RED queue size without ECN with 4 streams and drop probability equal to 1/50
  Figure 2: instantaneous and average RED queue size without ECN with 8 streams and drop probability equal to 1/50
  Figure 3: instantaneous and average RED queue size without ECN with 32 streams and drop probability equal to 1/50
  Figure 4: instantaneous and average RED queue size without ECN with 64 streams and drop probability equal to 1/50

• Comparison with and without ECN (64 streams, drop probability = 1/50):
  We can see that with ECN the overall burstiness is reduced, for example, the number of events where the instantaneous queue size is the range [140,160] packets is only 2 with ECN, while it's p to 6 without ECN (see Figure 5 and 6).
  Figure 5: instantaneous and average queue length for 64 ECN-capable streams and drop probabaility equal to 1/50
  Figure 6: instantaneous and average queue length for 64 non ECN-capable streams and drop probabaility equal to 1/50

• Comparison of 64 ECN-capable streams for different drop probabilities:
  From the comparison of the ECN effect in case of different drop probabilities (for a constant number of streams - 64 -) we can see for example that the number of itnervals during which the instantaneous/average queue length is non-zero is: always non-zero for drop probability = 1/50, 11 intervals for drop probability equal to 1/5 and only 9 intervals with drop probability equal to 1/2.
  Figure 6: instantaneous and average queue length for 64 ECN-capable streams with drop probability equal to 1/50 Figure 6: instantaneous and average queue length for 64 ECN-capable streams and drop probability equal to 1/5 Figure 6: instantaneous and average queue length for 64 ECN-capable streams and drop probabaility equal to 1/2

S.Alessandrini, T.Ferrari, March 22 2002