zaitcev.mee.nu

For a few short years, the backbone of computer networking in late USSR was Internet e-mail transmitted over UUCP. To hook MISS into it, I needed so-called "g protocol".[1] It provided basic flow control over a modem link with sliding windows. The counters were in packets, not bytes like in TCP, and up to 8 packets could be outstanding at a time before acked.

To be frank, it was a hell on Earth. Or almost, because it was fun after all, like climbing the Everest.

To begin, merely programming a sliding window protocol challenged my prowess. There's a million of corner cases, all the ack numbers had to be tracked, modulo-8 arithmetic comparisons were confusing. I struggled with basic correctness.

But the hell arrived when I faced interoperability. Our first counterpart at Demos was a Xenix of some kind. Its uucp descended from the earliest AT&T originals. I more or less made my implementation talk with that. But one day, Demos switched us over to a VAX with BSD 4.1.[2] That used an entirely different implementation, known as "pk". My code flat out didn't work with it, and it took a huge effort to fix it up. The final implementation that I had to support was UUPC, an MS-DOS program. By the time I had to talk to it, my code was tightened up considerably and UUPC only took some slight tweaks.

The amount of work it took to get g-protocol robust has framed my understanding of communicatio protocols for many years. It continues to amaze me that TCP even works at all, considering the diversity of networks that it must traverse, and above all the multitude of implementations. TCP is significantly more complex than g-protocol, even before we talk auxiliaries such as PMTU Discovery or extensions such as SACK. I didn't have to deal with sizing of the window and the Slow Start.

Unfortunately, they source code of my g-protocol implementation was lost when I changed jobs to MCST and didn't think to take it along.

Update: Michael Y. prompted me to date the events above. I happened to keep a printout of an e-mail, sent to avg at hq.demos.su in August 1990. That is when the g-protocol in MISS was complete enough that I started to look at RFC-822, and wanted a sample. But the work on g-protocol started back when .su TLD did not yet exist. I distinctly remember seeing e-mails addressed to ussr.com. So, a late 1989, I suppose.

[1] Aside of the g-protocol itself, I needed a few other things: a Hayes-compatible modem, and an ability to control it. It was not a trivial task to find hard currency in order to buy the 1200 baud modem, while working in a univirsity in a collapsing country. In addition, Mitra-225 cum ES-1011, as well as MISS had some trouble talking to the modem. The "channel" was barely capable of supporting full-dublex operations at 1200 baud with linked channel programs. Interrupts were expensive! The OS did not have any way to buffer input either. I ended writing a kernel driver that de-framed g-protocol and buffered 1 or 2 packets at a time. This allowed the userland to receive packets with MISS' equivalent of ioctl(2). All this was small potatoes, however.

[2] It was the famous kremvax.demos.su. My own node was r11740.phys.msu.su. R11 was yet another alias of Videoton's clone of Mitra, and 740 was the serial number of the CPU.

Next: Memoir 5.

Posted by: Pete Zaitcev at 08:56 PM | Comments (3) | Add Comment
Post contains 584 words, total size 3 kb.