RFC 54 An Official Protocol Proffering 18 June 1970
procedure has demonstrable limitations, but has the advantages that
it is more cleanly implementable and will support initial network
use. This is the only substantive change from the protocol already
agreed upon.
The new flow control mechanism requires the receiving host to
allocate buffer space for each connection and to notify the sending
host of how much space in bits is available. The sending host keeps
track of how much room is available and never sends more text than it
believes the receiving host can accept.
To implement this mechanism, the sending host keeps a counter
associated with each connection. The counter is initialized to zero,
increased by control commands sent from the receiving host, and
decremented by the text length of any message sent over the
connection. The sending host is prohibited from sending text longer
than the value of the counter, so the counter never goes below zero.
Ideally, the receiving host will allocate some buffer space as soon
as the connection is established. The amount allocated must never
exceed what the receiver can guarantee to accept. As text arrives,
it occupies the allocated buffer space. When the receiving process
absorbs the waiting text from the buffer, the NCP fires back a new
allocation of space for that connection. The NCP may allocate space
even if the receiving process has not absorbed waiting text if it
believes that extra buffer space is appropriate. Similarly, the NCP
may decide not to reallocate buffer space after the receiving process
makes it available.
The control command which allocates space is
ALL
This command is sent only from the receiving host to the sending
host.
This formulation of flow control obviates the RSM and SPD commands in
NWG/RFC #36, and the Host-to-Imp message type 10 and Imp-to-Host
message types 10 and 11 in the current revision of BBN Report 1822.
The obvious limitation in this scheme is that the receiving host is
not permitted to depend upon average buffer usage -- worse case is
always assumed. If only a few connections are open, it is unlikely
that there would be much savings. However, for more than a few
connections, average buffer usage will be much less than allocated
buffer space. We have looked at extensions of this protocol which
would include adaptive allocation, and we believe this to be
feasible. For the present this limited scheme seems best, and we
Crocker, Postel, Newkirk & Kraley