For those of you that can stroke a Sun, I have a bunch of scatter diagrams
showing some interesting characteristics of the ARPAnet, MILnet and the
gateways between them. For comparison, I also have scatter diagrams showing a
typical ARPAnet path in December, 1983. The diagrams can be FTPed from
UDEL2.UDEL.EDU (binary/image mode) and lit using the Sun screenload program.
Each diagram is about 40K octets in length and is stored in a file with BIT
extension. Most were made using fuzzballs either connected directly to the
ARPAnet or MILnet or behind a fast, lightly loaded gateway.
Each diagram shows delay versus length and was constructed using ICMP pings in
the manner described in RFC-889. These data should be considered only a sample
of Internet characteristics, so it is possible that collection of additional
data may reveal new surprises. Following are some brief comments, which should
be read with Sun in hand.
Files ISID.BIT and VENERA.BIT show a typical transcontinental path via
ARPAnet. The former reflects the path as of December 1983, while the latter
the path as of today. Regression lines are also shown on the diagrams. Note
the two-step delay characteristic for ISID, which was due to the ARPAnet
design at that time which used different allocation strategies for single and
multiple packet messages. The two-step characteristic is also apparent for
VENERA, but not as pronounced. Note the increased dispersion in the
contemporary data, which is hardly surprising to any of us.
Files ARPMIL.BIT, MILARP.BIT and ISIA.BIT show typical Internet paths between
hosts on ARPAnet and MILnet via an ARPAnet/MILnet gateway. ARPMIL shows an
ARPAnet path, MILARP a MILnet path and ISIA a combined path. The effect of
network load is clearly apparent when compared with VENERA. What bothers me
here is the huge dispersion at the lower packet lengths. A more clever routing
algorithm would show dispersion roughly proportional to length.
Files ARPNIC.BIT and MILNIC.BIT show typical ARPAnet (ARPNIC) and MILnet
(MILNIC) paths between east-coast hosts and the Network Information Center
host, which is connected directly to both nets. The effect of additional
trunking capacity on MILnet is obvious. Comparison with ARPMIL and MILARP
show that maybe that capacity is in the wrong place.
I have an extensive set of diagrams also for NSFnet and some of its
tributaries. It would be interesting to extend these measurements to other
paths, including SATNET, WIDEBAND and SURAN hosts. All it takes is a
convenient fuzzball and a vampire tap or alligator clips.
This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:37:00 GMT