Re: Mail Bridge Performance


Van Jacobson (van@lbl-csam.ARPA)
06 Mar 86 06:42:25 PST (Thu)


At least some of the pathetic Milnet - Arpanet performance should be
blamed on EGP. For example, EGP routes advertised on Milnet are all
via the east coast, usually bbn-milnet-gw. Routing all the west coast
Mil-Arpa traffic through Boston increases our transit delay, wastes
bandwidth on the transcontinental trunks and probably helps to saturate
the overloaded bbn bridge.

I just took some ping data between lbl-csam and ucb-arpa: Lbl-csam has
addresses 26.1.0.34 (milnet) and 128.3.0.24 (lbl-ether). Ucb-arpa has
addresses 10.0.0.78 (arpanet) and 128.32.0.4 (ucb-ether). This morning
around 3am, I pinged arpa from csam using both csam source addresses
and both arpa destination addresses. 100 packets were sent for each of
the four src/dest combinations then another 100 packets to each. The
two runs for each combination were separated by about 15 minutes but
had essentially identical statistics. Neither machine had active users
but there was sporadic inbound mail traffic. The results were (all
times in ms.):

                   Median Avg S.D. Min. Max %lost
  milnet - arpanet 195 361 360 100 2190 0
  milnet - ucbether 750 900 320 620 899 0
  lblether-arpanet 1060 1508 1043 500 4780 6
  lblether-ucbether 4430 4858 2757 1340 12160 7

The milnet-arpanet traffic used the correct route, millbl. The
milnet-ucb traffic used the lbl-csam EGP route, milbbn, outbound and
millbl inbound. The lbl-arpanet traffic used millbl outbound and the
ucb-arpa EGP route, milisi, inbound. The lbl-ucb traffic used milbbn
outbound and milisi inbound.

The "min" numbers scaled linearly and show a factor of 10 increase in
delay due to EGP. I don't understand why the avg./median numbers don't
scale linearly, why they show a factor of 20 increase in delay or why
packets were lost when routed through milisi but not when routed
through milbbn. Given the TCP retransmit time algorithms in 4bsd, I do
understand why all our telnet/ftp users are complaining: they could
walk to campus faster then their packets get there.

We've been suffering with EGP for more than a year and there still
isn't a west coast EGP server for Milnet. Perhaps adding one would
help users at both ends of the country. (This wouldn't be a fix but it
might be a cheap patch while we wait for butterfly bridges, smarter or
local EGP servers, or a replacement for EGP.)

 - Van Jacobson (van@lbl-csam.arpa)



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:36:04 GMT