Network Working Group Steve Crocker, Chairman
Request for Comments: 102 at BBN, Cambridge
NIC#5763 22, 23 February 1971
OUTPUT OF THE HOST/HOST PROTOCOL
GLITCH CLEANING COMMITTEE
At the NWG meeting in Urbana on 17-19 February 1971, a
committee was established to look at the Host/Host protocol
and see what changes were immediately desirable or necessary.
The committee is chaired by Steve Crocker, and has eight
other members:
Ray Tomlinson BBN (Tenex)
Jim White UCSB
Gary Grossman Illinois
Tom Barkalow Lincoln (TX2)
Will Crowther BBN (IMPs)
Bob Bressler MIT (Dynamic Modeling
Doug McKay IBM (Yorktown)
Dan Murphy BBN (Tenex)
A number of topics were discussed. On some of these topics, a
consensus was reached on whether or not to recommend a change, and if
so, what the change should be. On the remaining topics, specific
alternatives were proposed but no consensus was reached.
The committee will immediately canvas the network community and
gather reaction to its recommendations and the proposed alternatives.
The committee will then reconvene at UCLA on 8 March 1971 and decide
on final recommendations. Steve Crocker will then write Document #2.
This sequence is in lieu of the change procedure outlined in NWG/RFC
53.
Specific Recommendations
1. The ECO and ERP command should each be 8 bits long.
2. The ERR command should be 96 bits long.
3. Message Data Types should be eliminated. Third-level protocol
people may reinstate such a mechanism.
4. The Cease mechanism should be discontinued.
5. A new pair of one byte commands RST (reset) and RRP (reset reply)
should be added. The RST should be interpreted as a signal to
purge the NCP tables of any existing entries which arose from the
sending Host. The RRP command should be returned to acknowledge
receipt of the RST. The Host sending the RST may proceed after
receiving either a RST or a RRP in return. A RST may be returned
if the second Host comes up after the first Host.
6. Although it was suggested at the Urbana meeting that connections
should be full-duplex, the committee recommends against this
change.
7. Messages should be an integral number of bytes, and the number of
bytes and the byte size should be specified in each message. The
marking convention should be abandoned and the padding ignored.
The number of bytes in the message should be a 16-bit number
following the leader. The byte size should be in the next 8-bit
field. Two suggestions were generated for the starting point of
the text, and these are eXPlained in the next session.
For flow control purposes, the number of bits in a message is the
product of the number of bytes and the byte size. The leader and
other fixed format fields are not counted.
8. The problem of synchronizing the interrupt signal in a console
input stream was considered. We consider the console input
scanner as a process and note two reasonable implementations: it
may either read characters as fast as it can, looking for the
interrupt character and throwing away characters if there is no
room in the user process' input queue; it may read characters only
as fast as the user process can receive them, (or at least has
room for them).
The first implementation guarantees that the interrupt character
(e.g., control - C on the PDP-10 10/50) will always be acted on, but
requires that the using process interpret the output stream to detect
when it is sending too fast. The second implementation avoids
overrun but may not allow for sending an interrupt code. Note that
in the first case, allocation is alway renewed as soon as possible by
the console input interpreter; whereas in the second case, allocation
is renewed only as the result of acceptance of data by the user
process.
We decided that this is really a third-level protocol matter, viz,
use the INS to mean that a special code has been inserted into the
input stream. In conjunction with this, create the special code to
be put into the input sequence.
This special code would be network-wide and independent of the
particular interrupt character peculiar to the serving system. The
scheme for interrupting a serving process is that the using process
inserts the serving Host's interrupt sequence, followed by the
network special code, and also issue the INS.
UNRESOLVED ALTERNATIVES
1. Length of Control Messages
In accordance with other specifications, control messages should be
an integral number of 8-bit bytes, the length should be specified in
the byte count field, and control commands should not be split across
messages.
Unresolved was whether to specially limit the length of control
messages. The two choices are.
a) no special limit ( ~ 1000 bytes)
b) 120 bytes
2. Message Format
It was agreed to abandon marking and include the text length in the
form of a byte count and byte size. Unresolved was where to begin
the first byte of data. The two choices are:
a) have the first data byte begin after 72 bits of leader, byte
count, byte size and spacing. The message format would then be as
in the diagram:
<------------16------------>
__________________________
_ _ _ _ LEADER _ _ _ _
__________________________
BYTE COUNT
__________________________
BYTE SIZE--------->
_________________________
<--------------Beginning of first
_________________________ data byte
b) use the double physical transmission scheme presented in
NWG/RFC67. When sending a regular message, the Host would send a
leader, byte count and byte size and terminate transmission. The
second transmission would be the data.
At the receiving end, the IMP would transmit 64 bits of leader,
byte count, byte size and spacing, and stop transmission. The
next transmission would be only the data.
3. Allocation
With respect to the allocation mechanism embodied in the ALL, GVB and
RET commands, two alternatives were proposed:
a) make no change.
b) The flow control algorithm should be changed to keep track of
two quantities: messages and bits. The ALL, GVB, and RET commands
each have two data fields. The ALL command allocates a message
limit and a bit limit. The GVB command contains two fractions,
and the RET command returns both messages and bits. When sending
a message, the sending NCP decrements its message counter by 1 and
its bit counter by the text length of the message. The sending
NCP may not cause either of its counters to go negative. The
message counter would be 16 bits long.
[ This RFCwas put into machine readable form for entry ]
[ into the online RFCarchives by Gottfried Janik 02/98 ]