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

Re: Last Call: Security Architecture for the Internet Protocol to Proposed Standard

I fully agree. TCP window advertisements are supposed to show how much
buffer space the receiver has available. The receiver should not have
to artificially limit its window because of network considerations --
that's the function of congestion control on the transmitting end.

One problem we do have with the existing TCP congestion control
mechanism is that the sender will increase its congestion window all
the way up to the offered window as long as no packets are lost, even
if all those extra packets just pile up in a queue at the bottleneck
router. Various ad-hoc methods involving real-time bandwidth/delay
measurements have been tried to solve this problem, but they don't
seem to work really well because of measurement noise (competing
traffic, route changes, etc).

Signalling congestion by deliberately dropping packets when buffer
space is still available has always bothered me. There has to be a
better way. Besides, I'm of the opinion that there's just no
alternative to having lots and lots of router buffering, especially on
high speed/long delay links. If a TCP sends a packet into a
bottlenecked path with generous link buffering, the worst that happens
is that the packet waits in the queue for awhile. The link continues
to pass traffic at full capacity. But if TCP timidly shrinks its
congestion window to avoid packet drops in memory-starved routers,
that increases the chance of the link going idle even when the TCPs
have traffic to send.  Idle link time is gone forever.

Explicit congestion notification would help considerably.  This
doesn't necessarily require the definition of new bits in the IP and
TCP headers; you can actually do quite a lot with the oft-maligned
ICMP Source Quench message. I've found it quite effective in the
common case of a host on a private leaf Ethernet sending data through
a dialup router to the outside world. It certainly reacts more quickly
than a mechanism that requires a far-end turnaround of the congestion
bit.

Phil
References: