[Text version]

Network Working Group Request for Comments: 50

E. Harslen J. Heafner RANL 4/30/70

We find the Meyer proposal (Note #46) to be the most acceptable to dare, for exactly the reasons that he enumerates; viz., simple, suffices for most planned uses of the Network, easy to implement, can be extended. It does not encompass everything that has been suggested recently, however, we do agree with the items that are proposed and we feel that the missing features are probably not worth doing battle over and thus delaying the specification.

We make the following comments on the seven issues rasied in Note #47.

   1)  We agree with Steve that dynamic reconnection will later
       be required for more sophisticated uses of the Network.
       We also agree with the Project MAC people that it
       unnecessary initially.  A better job can be done of dynamic
       reconnection given some Network experience and the specific
       needs of its use.

2) INT is easy to implement and serves a useful purpose.

   3)  We favor including a sub-field for instance tag identifier.
       We see the need for both cases; a) where multiple processes
       should appear indistinguishable, and b) where a given
       user owning multiple processes must distinguish among
       them.  Those program parts that should not distinguish
       among processes should simply ignore the instance tag.
       Tom's suggestion to use part of the user number sub-field
       merely reduces the combined length of sub-fields from 32
       bits to 24 bits; the problem remains.

   4)  We disagree with both Steve and MAC in that no special
       structure should be imposed on the data transmitted.  We
       prefer the "message data type" mentioned by E. I. Ancona,
       Note #42, page 1.  An example of its use was cited in
       Note #39, page 2, transmit vs broadcast.

       support adopting one in the beginning, and in particular
       ASCII.  We have observed that most sites have previously
       suggested ASCII.  Is there anyone who objects?

   5)  Word boundary alignment is more attractive than double
       padding.

   6)  Steve's suggestion of short-term queueing of RFCs is
       acceptable as an option.

7) We support the UCC in Note #46 for three principle reasons:

       a)  In general the user should not know the remote socket
           code of the process to whom he wishes to communicate.

       b)  The additional duplex connection can provide some
           superfisory control over process behavior, possibly
           in conjunction with the interrupt procedure.

       c)  Most of the other proposed methods demand queueing.

      We think there must be a standard UCC, yet we encourage
      parallel experimental UCCs.

We make two additional comments on Note #46 that were not reiterated in Note #47.

BLK and RSM are more straightforward than previous suggestions and they do not deny multiplexing over a given link. With regard to the use of links, we refer to an example given by Bob Kahn where an intermediate IMP goes down and eats some's RFNM. This should not necessitate reconnection.

In Note #46, page 6, the statement that the UCC has the ability to close connections to a dead process is installation dependent. In our particular case the NCP is notified directly of process failure due to the particular software interface through which all processea, including NCP, must communicate.

JFH:hs

       [ This RFC was put into machine readable form for entry ]
          [ into the online RFC archives by Gary Okada 7/97 ]