Cell Relay Archive

[murray@pa.dec.com (Hal Murray)]

> > Yes, it's complex.  Our implementation doesn't scale to WANs.
> 
> Complex?  I always thought that QFC (credit based flow control)
> was incredibly simple.  Certainly it's orders of magnitude
> simpler than ABR.

The stuff I'm talking about is the grandfather of QFC.  We put the
credits in the header (stole some bits) which made things much simpler
since the switch logic never had to look inside the body of a cell.

Yes, it's simple on one level.  I can explain it at a whiteboard
in a few minutes and anybody will probably agree that it will do
what I claim.  But it didn't seem simple to implement.  Maybe
the complexity was ATM in general rather than the flow control issues.

 
> Which flow control, QFC or ABR?  Also, I think the suspicions
> apply mostly to the WAN case (for ABR).  It seems reasonable,
> intuitively, that embedding another servo loop below TCP is
> fine when the time constants of that lower loop are substantially
> smaller than those of the upper one -- as they can be for LANs.

I'm talking about the grandfather of QFC.  I've never worked with
any rate based gear.


I like your time-constant line of reasoning.  I'm not quite sure what
it means in this case.  The time constant for flow control is roughly
the round trip time.  (If the congestion is in the other end, you won't
see any change untill it gets to you.)  What's the time scale for TCP?
The round trip time or the minimum retransmit time?  If it's the round
trip time, then it's not significantly different from the flow control
time.

As far as I can tell, TCP works fine if you never drop packets.  That
may depend on packets not getting out of order.

-- 
These are my opinions, not necessarily my employers.