Mills@UDEL.EDU (03/24/88)
Folks, You might be interested in the following summary of data on NSFNET Backbone Fuzzball performance, as well as a comparison with the ARPANET/MILNET gateways, which operate in a similar state of massive traffic insult. These data update those reported in the SIGCOMM 87 paper and include the period since the source-quench policy described previously (and in a paper just submitted) was implemented. In July 1987 after the Fuzzball selective-preemption policy was installed, but before the source-quench policy was implemented, the throughput per trunk ranged from 0.3 to 4.0 packets per second, with a mean of 2.0. At that time the total trunk throughput was 31 pps with drop rate due all causes of .09 percent. In March 1988, several months after quench was installed, the trunk throughput ranges from 2.6 to 9.2 packets per second, with a mean of 4.7. At this time the total trunk throughput is 71 pps with drop rate 0.48 percent and with 0.27 percent of all trunk packets resulting in a quench. The following table shows the performance of the NSFNET Backbone Fuzzballs for periods ending on 21 March. These numbers include Ethernets as well as all 56-kbps trunks. Node Mpkt UpHr PPS Drop Quench ---------------------------------------------- 1 3.49 100 9.75 0.17 0.04 2 18.48 260 19.72 0.38 0.39 3 6.33 102 17.33 0.16 0.17 4 15.08 262 15.99 0.31 0.45 5 19.36 266 20.20 0.85 0.18 6 2.97 48 17.35 0.17 0.14 7 7.02 262 7.44 0.71 0.04 ---------------------------------------------- Total 72.74 1299 15.55 0.34 0.18 The "Mpkt" column shows the aggregate throughput in megapackets for all output queues, including serial lines and Ethernet. The "UpHr" column shows the aggregation interval in hours. The "PPS" column down through the "Total" row shows the resulting throughput, which is the "Mpkt" column divided by the "UpHr" column adjusted to the proper units. The "Drop" and "Quench" columns show the percentage of packets dropped and quenched respectively. The value shown in the "Total" row for these columns is the average of the column itself. The existing NSFNET Backbone clearly meets the performance objective of less than one percent drop rate. For comparison the following table shows the performance of the ARPANET/MILNET gateways for the week ending 21 March. So far as can be determined, each gateway is connected to two 56-kbps data paths. ID Mpkt UpHr PPS Drop ------------------------------------- 1 4.83 144 9.32 7.26 2 6.15 144 11.86 8.18 3 7.06 146 13.48 7.40 4 7.03 139 14.08 12.87 5 3.14 145 6.00 0.83 6 3.75 109 9.54 3.23 7 5.07 146 9.66 2.85 8 2.76 129 5.95 3.65 ------------------------------------- Total 39.79 1101 10.04 5.78 As evident from these figures, the NSFNET Backbone Fuzzballs carry a throughput over fifty percent greater per node than the ARPANET/MILNET gateways with a drop rate of over ninety percent less. Note that this comparison may not be fair in two ways: first, the ARPANET/MILNET gateways are connected to networks, not trunks, which can have large dispersive delays; second, the NSFNET Backbone Fuzzballs are connected to Ethernets, which provide no insulation against unruly traffic generators. From measurements made last July and reported in the SIGCOMM paper last year, the selective-preemption policy made a whale of a difference. The case for the source-quench policy installed recently is less clear, although there is recent evidence that it is in fact effective for those hosts that respond to quench messages. However, even if the crafted policies and Fuzzball implementations may be suboptimal and change next Monday, the data above should be convincing beyond doubt that fairness policies and queue disciplines similar to these will be necessary for future generations of connectionless packet switches and gateways. Dave