Request for Comments: 1018 BBN Labs
August 1987
Some Comments on SQuID
Status of this Memo
This memo is a discussion of some of the ideas eXPressed in RFC-1016
on Source Quench. This memo introdUCes the distinction of the cause
of congestion in a gateway between the effects of "Funneling" and
"Mismatch". It is offered in the same spirit as RFC-1016; to
stimulate discussion. The opinions offered are personal, not
corporate, opinions. Distribution of this memo is unlimited.
Discussion
It appears to me that there are at least two qualitatively different
types of congestion which may occur at Internet gateways. One form
of congestion is the result of the merger of several independent data
streams from diverse sources at a common point in their communication
path. I'll refer to this as "Funneling". The architecture of the
Internet (apparently) assumes that traffic flows are bursty and
asynchronous; therefore congestion which occurs at the result of
Funneling will typically be the result of "bad luck" as several
independent bursts happen to arrive at a common point simultaneously.
It is expected that Funneling congestion will be short-lived, just as
individual bursts are. I don't claim that any such assumptions are
documented or formalized; nevertheless I got a clear sense of this
class of assumptions both from reading the protocol documentation and
from personal recollections of long-past design meetings.
A second form of Internet congestion may arise during a prolonged
(non-bursty) data transfer between hosts when the resulting traffic
must pass through a node connecting two communications subsystems
with significantly different throughput rates. I'll refer to this as
"Mismatching". By contrast with Funneling, Mismatching can be caused
by the innocent action of a single host, is highly repeatable
(definitely not just "bad luck"), and will be long-lived.
RFC- 1016 discusses two interrelated strategies; one for when to send
a SQ, and a second for what to do when an SQ is received. There is
also a discussion of some experiments, which deal almost exclusively
with Mismatching congestion. (I understand that the simulation can
generate multiple flows, but these simply further increase the degree
of Mismatch; the flow under study is long-lived by design.) It seems
to me that the strategy RFC- 1016 proposes for sending SQ's, based on
queue length, may be appropriate for Funneling Congestion, but
inappropriate for Mismatch congestion, as discussed below. The host
behavior proposed in RFC- 1016 may be appropriate for both cases.
Since we assume that Funneling congestion is the result of short-
lived phenomena, it is appropriate for gateways which are the sites
of this congestion to attempt to smooth it without excessive control
actions. This is the basis for the "hint" in the ICMP specification
that maybe an SQ should be sent only when a datagram is dropped. It
is the basis for the idea in RFC- 1016 that a gateway should be slow
to cry "congestion" (SQK = 70% of queue space filed), even if
persistent in attempting to eliminate it (SQLW = 50% of queue space
filled). Since Funneling congestion is the result of the actions of
multiple senders, the growth of internal queues is the only
reasonable place to watch for its existence or measure its effects.
Mismatch congestion, on the other hand, is the result of incorrect or
inadequate information about available transmission bandwidth on the
part of a single sender. The sending host has available to it
information about destination host capacity (TCP window size and ACK
rate) and about local link capacity (from the hardware/software
interface to the directly-connected network), but no real information
about the capacity of the Internet path. As noted in RFC-1016, hosts
can oBTain the best throughput if their datagrams are never dropped,
and the probability of dropped datagrams is minimized when hosts send
at the appropriate steady-state rate (no "bunching"). Therefore, it
is a disservice to a host which is the source of Mismatch congestion
to wait a "long" time before taking control action. It would be
preferable to provide immediate feedback, via SQ's, to the host as
soon as datagrams with too short an inter-arrival time begin to
arrive. The sending host could then immediately (begin to) adjust
its behavior for the indicated destination.
There are, of course, many implementation issues which would need to
be addressed in order to implement the type of SQ-sending behavior
suggested here. Perhaps, though, they are not as severe as they
might appear. Two specific issues and possible solutions, are:
1. How should a gateway differentiate between Funneling and
mismatch congestion? Perhaps whenever there are more than q"
items on an output queue to a slower subnet which have been
received from a faster subnet, then look to see if any h" of them
have the same source. It so assume Mismatch and send an SQ to
that source. The "q" test might be implemented by a small set of
counters which are incremented when a packet is placed on an
output queue and decremented when a packet is sent. The search
for a common source might require more cycles but be performed
less often. The value of "q" would have to be small enough to
give an early warning, but bigger than a small multiple of "h".
The value of "h" would have to be big enough to avoid triggering
on common cases of source datagram fragmentation by an
intermediate gateway.
2. How can a gateway determine which subnets are "slower" and
faster", as well as appropriate inter-arrival times? There may be
lots of clever ways for a gateway to measure the dynamic bandwidth
of its directly-connected subnets. However, I'm more in favor of
starting with configuration parameters related to the known (at
installation time) general characteristics of subnet types (e.g.
Ethernet is 10Mbps, ARPANET is 50 Kbps, SATNET is 100 Kbps, etc).
This sort of approximation is quite adequate for determining which
subnet is faster, or what inter-arrival time is appropriate for
packets being routed to a slower subnet.
Summary
Funneling congestion and Mismatch congestion are qualitatively
different, and it would not be surprising if different SQ-sending
strategies were best for dealing with them. RFC- 1016 suggests a
specific SQ-sending strategy which may be inappropriate for dealing
with Mismatch congestion. This RFCsuggests guidelines for an
additional SQ-sending strategy for dealing with Mismatch. Hosts
implementing the SQuID algorithm of RFC-1016 should be expected to
achieve better performance if they received SQ's sent according to
either or both of these strategies. However, all these ideas are
still only in half-baked form; real engineering is clearly needed.