[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[ih] Secret precedence schemes back then

Wow, you have a good memory.  I still have a copy of the TCP code that I
wrote (first Unix TCP, on a PDP-11/40).  I'll have to look and see what
it did with Source Quench.

When I first got that TCP running, it didn't have enough throughput to
keep a model 33 TTY busy.  I think it achieved the blinding speed of
about 4 bytes/second!  It turned out that 99% of the time was spent in
the Unix kernel.  So, that's when I became a Unix kernel hacker, to fix
the O/S so the TCP could run.  Talk about hardware limitations - the
Unix system had 32 KB (yes, K) of memory.  Adding code required first
finding some error message to shorten; every character gone freed up a
byte for code, and made the kernel output logs slightly more
inscrutable.  Needless to say, there wasn't a whole lot of flexibility
to add much sophistication.

Yes, it was an interesting time.  The ELF implementation preceded me by
a bit.  As I remember, ELF made a good platform for a research project -
where performance wasn't much of an issue.  Part of the reason for
moving the project internally within BBN was to get a more "operational"
perspective.  So it moved into the group where I was, along with all of
the Arpanet crew.  Much of the innards of the ensuing gateways were
inspired, or at least influenced, by the Arpanet experience - the
gateway crew was literally down the hall from the Arpanet crew.


On Wed, 2009-01-28 at 15:24 +0000, David Mills wrote:
> Jack,
> The fuzzies had two throttles, one the current window size, the other 
> the number of outstanding packets. That number was monitored not to 
> exceed eight in view of the limited number of packet buffers in Ginny's 
> gateways. When a source quench arrived, that number was reduced on the 
> expectaion that Ginny or the fuzzies would hurl a source quench. So far 
> as I remember, Ginny never did, but the fuzzaies did.
> It was an interesting time. Ginny was stuck with a real resource hog 
> (Elf?) with a maximum throughtput of 10 pkt/s, but the fuzzies had much 
> more memory and a throughput of 300 pkt/s.
> Jack Haverty wrote:
> >On Tue, 2009-01-27 at 21:14 +0000, David Mills wrote:
> >  
> >
> >>but so far as I knew, nobody but the Fuzzballs 
> >>actuall responded to source quencn. 
> >>    
> >>
> >
> >Errr, ummm, well...depends on what you mean by "respond".
> >
> >Since Source Quench was sent by a receiver when it had gotten so
> >overwhelmed that it threw away your packet, the obvious response from
> >the Sender was to re-send the packet immediately, since you had just
> >been told that it had been discarded.
> >
> >I can't remember exactly what the various TCP implementations did that I
> >was involved in.  Or I could take the fifth amendment...
> >
> >Of course, the spec might have said something a bit different about what
> >a well-behaved TCP should do when you received a Source Quench.  But I
> >don't recall there ever being any "certification" or the like that any
> >particular implementation was behaving correctly.  As I remember, the
> >spec wasn't very specific.  E.G., If you send one packet and get a
> >Source Quench back, what does it mean to "throttle back".
> >
> >And of course I can't remember whether the core gateways put such
> >"control" traffic at the front of the queue (it's important stuff!), or
> >discarded it (gaaak, more whining and noise from that complainer
> >host...).  Or whether they looked at all.  Possibly all three schemes,
> >over time.
> >
> >Of course, this whole Internet thing was a research project that was
> >supposed to go away and be replaced by the "real" system using ISO and
> >CCITT technology.
> >
> >Good thing it's been 23+ years Dave!
> >
> >/Jack
> >
> >
> >  
> >