Summary

  • D. Smallberg's September 1983 RFC 876 reported tests of 483 hosts from an ARPANET machine and an ISI-Net machine. Of 283 hosts advertised in the NIC table as supporting SMTP, 49 allowed a connection from neither test site. Another 51 had no SMTP claim in the table but did allow a connection from at least one test site.
  • A connection was only the first layer. Of 285 hosts reached from the ARPANET test machine, 69 gave a positive VRFY postmaster reply, 162 immediately rejected a made-up recipient, and 115 acknowledged that bad recipient at the RCPT TO stage. Protocol replies and access differed by command and by the tester's location.

Someone had written “SMTP” next to a host in the 1983 NIC table. That sounded useful if you had mail to send. D. Smallberg of ISI tried opening connections and discovered that a directory entry could be more optimistic than the machine behind it. The surprise also ran the other way: some hosts without an SMTP claim answered. RFC 876 preserved the exercise as a dated survey, with the host list and individual observations in its appendix. The paper label and the open socket had started an argument.

Smallberg tested from two places between May and August 1983: ISI-VAXA on the ARPANET and ISI-MOE on ISI-Net. The starting list was the NIC host table of 17 August; TACs and echo hosts were left out. The 483 tested entries yielded a blunt first cut. The table claimed SMTP for 283 hosts. Forty-nine of those 283, or 17%, permitted a connection from neither test machine. Fifty-one entries lacking that claim permitted a connection from at least one. A host table was a statement about a host; a connection attempt was a statement about a particular path at a particular time.

The two starting points did not see the same network. ISI-VAXA opened connections to 285 hosts. ISI-MOE opened 170, and every one of those 170 was also reachable from ISI-VAXA. The remaining 115 of the 285 connectable hosts answered the ARPANET tester alone. That is a useful result precisely because it keeps the vantage points attached. A mail service available from one network could look absent from another. One successful socket call did not settle the route from every sender.

Then the survey asked what the server actually did with mail commands. Smallberg tried VRFY postmaster; 69 of the 285 connectable hosts replied positively. That figure is tempting to turn into a roll call of helpful mail administrators, but the SMTP specification treated VRFY as an optional command. The survey also sent RCPT TO for postmaster and for the fabricated account jqkxwzvb. These were different tests. A server could reject the fictional user immediately, accept a postmaster recipient at that point, or postpone its decision about the fictional account. The reply at RCPT TO marked one stage of a mail transaction.

The bad address produced an especially instructive split. Among 285 connectable hosts, 162 immediately answered that the invented user was unknown. Another 115 gave the RCPT TO command a positive acknowledgement; eight were never up during that part of the test. Of the 162 immediate rejecters, 121 accepted postmaster as a recipient at the same stage. Smallberg reported those 121 as 75% of the immediate rejecters, or 42% of all 285 connectable hosts. The denominator changes with the question. So does the thing being measured: opening a connection, verifying a name, accepting a recipient command and delivering a finished message sit at different points in the exchange.

Smallberg also saw protocol manners that would make a standards author reach for a red pen. Around 40 hosts had at least one of three significant problems. Some answered RSET or NOOP with code 200, which RFC 821 never assigned as a legal reply, instead of 250. Some answered VRFY before MAIL with 554 Nested MAIL, even though the SMTP specification allowed VRFY anywhere in the session. Others sent a nondelivery message after RSET canceled a transaction before any message body had been collected. Smallberg noted that RFC 821 did not explicitly ban that last behavior; it sat badly with the idea of resetting a transaction. These findings concerned what happened after the door opened.

The survey's value lies in keeping its measurements separate. The host table described intended support. Two test machines observed connection access. SMTP commands probed behavior once inside. The appendix tied those observations to named hosts at the time of the test, while the summary exposed the larger mismatch. It was a snapshot of working implementations, with bugs reported and many fixed during the survey period. Each repair could change the next visit. The result was a picture of an Internet that had already learned a lasting lesson: an address book cannot run a mail server for you.