Network Working Group                                    M. A. Padlipsky
Request for Comments: 928                                    Mitre Corp.
                                                           December 1984

               INTRODUCTION TO PROPOSED DOD STANDARD H-FP


Status Of This Memo

   This RFC suggests a proposed protocol for the ARPA-Internet
   community, and requests discussion and suggestions for improvements.
   Distribution of this memo is unlimited.

Important Prefatory Note

   The broad outline of the Host-Front End Protocol introduced here and
   described in RFC 929 is the result of the deliberations of a number
   of experienced H-FP designers, who sat as a committee of the DoD
   Protocol Standards Technical Panel under the author's chairmanship.
   The particular protocol to be described is, however, the result of
   the deliberations of a small, ad hoc group, who sat as a de facto
   subcommittee of the H-FP committee, also under the author's
   chairmanship. The protocol, then, follows the consensus of the full
   group as to what the new H-FP should "look like," but has not
   benefitted from painstaking study by a large number of experienced
   H-FP designers and implementers.  (It has been looked at before
   release as an RFC by several of them, though.) Even if that were not
   the case, it would still be the intent of the designers that the
   protocol be subjected to multiple test implementations and probable
   iteration before being agreed upon as any sort of "standard".
   Therefore, the first order of business is to declare that THIS IS A
   PROPOSAL, NOT A FINAL STANDARD, and the second order of business is
   to request that any readers of these documents who are able to do
   test implementations (a) do so and (b) coordinate their efforts with
   the author (617-271-2978 or Padlipsky@USC-ISI.ARPA.).

Historical/Philosophical Context

   Late in May of 1971, the author was presenting a status report on
   whether the Multics ARPANET implementation would be ready by the
   July 1 deadline declared by the sponsor earlier that month.  Some
   controversy developed over the fact that the Multics "NCP" (Network
   Control Program--actually a blanket term covering the Host-Host and
   Host-IMP protocol interpreters) did not queue requests for
   connections.  As the specification explicitly declared the topic to
   be one of implementors' choice, the author attempted to avoid the
   argument by asking the interrogator what he was up to these days.
   The answer was, "Oh, I'm working on the High-Speed Modular IMP now"
   (later the Pluribus IMP).  And the proverbial coin dropped:  The
   author replied, "I've got a great idea.  Now that we've got some
   space to program in the IMP, why don't we separate out most of the


Padlipsky                                                       [Page 1]


RFC 928                                                    December 1984
Introduction to H-FP


   NCP and do it outboard: the only thing that really matters in the
   Host is associating sockets with processes, and if we had common
   implementations of all the bit-diddling stuff in the IMPs, we
   wouldn't have disputes over the interpretation of the spec and we'd
   also save a lot of Host CPU cycles!"

   As far as the author knows, that incident was the beginning of what
   came to be called "Network Front-Ends" and, more recently, "Outboard
   Processing Environments."  (The name change, by the way, was
   motivated by a desire to prevent further confusion between NETWORK
   Front Ends--always conceived of as distributed processing mechanisms
   for the offloading of intercomputer networking protocols from
   Hosts--and traditional communications front-ends, which have no
   connotation of bearing protocol interpreters invokable by Host-side
   programs.)  At least, the idea was original to him and he later was a
   principal designer and the primary author of the first Host-Front End
   Protocol.  So, on the one hand, the present document might be marred
   for some readers by undertones of parental pride, but on the other
   hand, if you like primary sources....

   The evolution of the outboard processing idea has been dealt with
   elsewhere [1]. For present purposes, it should suffice to observe
   that some half-a-dozen implementors of "NFE's" of various sorts are
   known to the author to have met with success.  The topic of why use
   an explicit protocol in the first place (as opposed to emulating a
   device, or devices, already known to the Host/operating system)
   deserves a word or two here, however.  ([2] deals with it in more
   general terms.)  The crucial consideration is that in the general
   case you wind up "not doing real networking" if you attach a Host to
   a network by known device emulation, where real networking is taken
   to mean what has been called "resource sharing" in the ARPANET
   literature, and what appears to be dubbed "open system
   interconnection" in the ISO literature: Operating systems' built-in
   assumptions about known devices--whether terminals, terminal
   controllers, or RJE stations--tend to get in the way of the sort of
   process-process and eventually procedure-procedure communications
   that serve as the basis for applications more interesting than simple
   remote login.  To those unfamiliar with the outboard processing
   approach, the premise that the way to attach is via an explicit
   protocol may be difficult to accept, but to those who have done it,
   it makes almost perfect sense.

   To those, by the way, who have worked in intercomputer networking
   from the perspective of inboard (Host-side) implementations of
   protocol suites, the outboard processing idea often seems to lead to
   less than optimal results, especially as to maximizing throughput.
   And it is difficult to argue that if a given Host were well and truly


Padlipsky                                                       [Page 2]


RFC 928                                                    December 1984
Introduction to H-FP


   fine-tuned to "do networking" the insertion of an extra processor
   could somehow lead to better networking.  However, for Hosts where
   conservation of CPU cycles is an issue, or even where memory is
   scarce (i.e., where it's desirable to conserve the resources being
   shared), outboarding is clearly the way to go.  For that matter,
   viewing outboard processing aright (as a form of distributed
   processing) it can be argued that even for extremely powerful
   "intelligent work stations"/"personal computers" which have the
   resources to spare it still makes sense to outboard in order not to
   have to do new implementations of entire protocol suites for each new
   such system--always assuming, of course, that the Host-Front End
   protocol in play is noticeably less complex than the offloaded
   protocols.

   None of this is meant to imply that outboard processing is the ONLY
   way to do intercomputer networking, of course.  It is, however, meant
   to suggest that outboard processing can be advantageous in a number
   of contexts.  Indeed, given the joint advents of microprocessors and
   Local Area Networks, a generic bus interface unit which also plays
   the role of a NFE (that is, is an Outboard Processing Environment)
   even allows for the original intent of "offloading to the IMP" to be
   realized, so that a free-standing, possibly fairly expensive NFE need
   not be interposed between Host and net.  Note, by the way, that
   nothing in the OPE approach requires that ALL Hosts employ OPEs. That
   is, the only protocols "seen" beyond the Comm Subnet Processor are
   the common intercomputer networking protocols (e.g., all DDN IMPs see
   and read IP datagrams). H-FP is strictly a matter between a Host and
   its OPE.

   It is also important to be aware that, given the advent of several
   different suites of protocols in the networking world, it might well
   be the case that the only reasonable way to achieve
   "interoperability" might well be to use a suitable H-FP (such as the
   one to be presented in the companion RFC) and an Outboard Processing
   Environment which is capable of parallel invocation of protcol suites
   (with the choice of suite for a given connection being dependent, of
   course, on the native suite of the desired target Host and/or
   application).

   The unquestionable advantages, then, of the approach, based on ten or
   more years of experience and analysis, would seem to be as
   follows--always recalling the assumption that the work to implement
   and execute the H-FP in play is small compared to the full protocol
   suite in question:  As noted, common implementation of a protocol
   suite has the automatic advantage of mutual consistency; further,
   particularly in the DOD context, it's far easier to procure common



Padlipsky                                                       [Page 3]


RFC 928                                                    December 1984
Introduction to H-FP


   implementations of standard protocols than to procure different ones
   on a per-Host type basis.  Also as noted, if the resources to be
   shared are viewed as being the participating Hosts'

   CPU cycles and memories, these resources are conserved by doing  as
   much as possible of the networking protocols in an OPE rather than in
   the mainframe.  Another, less evident advantage is that having an OPE
   effectively insulates a Host against changes in the
   outboarded/offloaded protocols--or even changes of the protocols,
   should the nascent international protocol standards ever mature
   sufficiently to supplant the in-place DOD standards.  (That is, given
   an abstract enough interface--in the spirit of the Principle of
   Layering--a Host could, for example, go from doing TCP as its
   "Host-Host" protocol to, say, ECMA Class 4 as its "Transport"
   protocol without taking any particular cognizance of the change,
   however unattractive such a change would be to advocates of the
   APRANET Reference Model such as the author. See [3] for more on the
   implied "Reference Model" issues.) Finally, although a few rather
   specialized points could also be adduced, it should be noted that for
   network security architectures which are predicated on the ability to
   control all means of egress from and ingress to "the net", uniform
   use of OPEs is clearly desirable.

   If we can stipulate that an OPE is/can be a good thing, then the
   remaining problem is just what the protocol interpreted by a Host and
   its OPE ought to be, once it's observed that a standard protocol is
   desirable in order to allow for as much commonality as possible among
   Host-side interpreters of the protocol.  That is, we envision the
   evolution of paradigmatic H-FP PIs which can more or less
   straightforwardly be integrated with  various operating systems, on
   the one hand, and the ability simply to transplant an H-FP PI from
   one instance of a given operating system to other instances of the
   same system, much as is currently being attempted in the DODIIS NFE
   program.  Again, the major motivation in the DOD context is the
   minimizing of procurement problems.

Technical Context

   As noted, some half-a-dozen Host-Front End protocols have been seen
   by the author.  Indeed, in December of 1982, a meeting was convened
   to allow the developers of those H-FPs to compare their experiences,
   with an eye to coming up with a proposal for a DOD standard H-FP;
   this paper is a direct result of that meeting.  In the current
   section, we present the consensus of the meeting as to the broad
   outline of the protocol; in the accompanying document, the current
   version of the proposed protocol will be presented, as detailed by
   the author and Richard Mandell and Joel Lilienkamp (both of SDC).


Padlipsky                                                       [Page 4]


RFC 928                                                    December 1984
Introduction to H-FP


   Note, by the way, that in some sense we should probably have changed
   the name from H-FP to H-OPEP (or something), but the habit of saying
   "H-FP" seems too deeply engrained, despite the fact that it does seem
   worthwhile to stop saying "NFE" and start saying "OPE."  (Besides,
   H-OPEP looks rather silly.)

   A final preliminary:  all the designers and implementors of H-FPs
   present at the December meeting concurred that the true test of any
   protocol is how well it implements.  Therefore, until several
   implementations of the "new" protocol have been performed and
   assessed, it must be understood that the proposed protocol is
   precisely that:  a proposal, not a standard.

   Not too surprisingly, the first point on which consensus was reached
   is that there are three separable aspects (or "layers") to an H-FP:
   At bottom, there must be some physical means for conveying bits from
   Host to OPE and from OPE to Host.  As it has always been a premise of
   outboard processing that the Host's convenience is paramount, just
   what this physical layer is can vary:  typically, a bit-serial
   interface is customary, but parallel/DMA interfaces, if available for
   the Host and interfaceable to a given OPE, are fair game.  (So would
   teleporting the bits be, for that matter.)

   In the middle, there must be a layer to manage the multiplexing of
   network "connections" and the control of the flow between Host and
   OPE.  If we agree to call the lowest layer the Link and the middle
   layer the Channel, one thing which must be noted is that between the
   two of them, the Link and Channel layers must be responsible for
   reliably conveying the bits between Host and OPE. After all, an OPE'd
   Host should not be "weaker" than one with an inboard implementation
   of a robust Host-Host protocol such as TCP.  It should be noted that
   any Host which "comes with" a suitable implementation of the X.25
   interface protocol (where the definition of "suitable" is rather too
   complex to deal with here) could, given an OPE conditioned to accept
   it, quite cheerfully satisfy the requirements of the lower two
   layers. This is not to say that X.25 "is" the mechanization of H-FP's
   Link and Channel layers, however; merely that it could be used.  The
   protocol spec itself will detail an alternative, less cumbersome
   channel layer for Hosts which don't have or want X.25.

   The top layer of H-FP is the most important:  we refer to it as the
   Command layer.  Here is where the peer H-FP modules in a given Host
   and OPE communicate with each other. Indeed, the segregation of JUST
   multiplexing and flow control (plus reliability) into the Channel
   Layer is done--in addition to making it easier for Hosts that possess
   preexisting software/hardware which could be turned to the
   purpose--so as to clarify "what the H-FP is":  it's the commands and


Padlipsky                                                       [Page 5]


RFC 928                                                    December 1984
Introduction to H-FP


   responses of the Command layer wherewith the Host's processes are
   able to manipulate the outboard implementations of the members of a
   protocol suite. The use of the phrase "commands and responses" is
   rather significant, as it happens. For in the protocol to be proposed
   for DOD standardization, unlike all but one of its predecessors,
   binary encoded "headers" are not employed; rather, the H-FP commands
   are indeed ASCII strings, and the responses (following the practice
   of ARPANET FTP) ASCII-encoded numbers.

   There are various reasons for this departure, which initially stemmed
   from a desire to have the same NFE be usable for terminal traffic as
   well as Host offloading, but the one that seemed to dominate when
   consensus was arrived on it as the basis for the new standard is that
   it is very much in the original spirit of H-FP.  That is, if you want
   to "make things as easy as possible for the Host", it makes a great
   deal of sense to offload in a fashion that only requires some sort of
   scenario or script ("exec-com"/"command file"/"shell command" are
   approximations on some systems) in the Host, rather than requiring a
   program, possibly of more complexity than we would like. This is not
   to say that we envision all--or even most--Hosts will take the
   scenario approach to H-FP mechanization, but rather that the command
   orientation chosen allows for the possibility. (It would be useful to
   recall that the Channel layer does all the necessary
   multiplexing/demultiplexing, so that each channel's  metaphorical
   state machine--at least on the Host side--really has very little to
   worry about other than "doing its thing.")

   It should be noted that the proposed protocol provides a mechanism
   for offloading "all" protocols.  That is, although most "first
   generation NFEs" only handled ARPANET Reference Model Layers II and I
   (Host-Host and Network Interface--approximately ISO levels 4-1, with
   some of L5's functionality included when it comes to service
   identifications being handled via Well-Known Sockets in L II), it is
   assumed that OPEs will be evolved to handle L III offloading as well
   (ISO 5-7).  Indeed, it should also be noted that what is being
   addressed here is "the protocol", not "the" OPE.  More will be said
   on this topic below, and in the protocol spec itself, but it is
   important to realize from the outset that the H-FP being proposed is
   intended to be implementable by any number of OPE suppliers/vendors,
   so "an" OPE may or may not choose to implement, say, a given file
   transfer protocol, but provided it says so in proper H-FP terms and
   does offload some other protocols it's still an OPE in our sense of
   the term. (Cf. "Issues" and "Non-Issues", below.)






Padlipsky                                                       [Page 6]


RFC 928                                                    December 1984
Introduction to H-FP


Issues

   The following items are either in some sense still open issues or
   bear special emphasis:

   Command Approach

      The most striking feature of the new H-FP, especially to those who
      have seen older H-FPs, is the decision to employ
      character-oriented commands rather than the more conventional
      binary-oriented headers at the Command Layer.  As noted, the
      primary motivation was the report that the approach worked well
      when it was employed in an H-FP for the Platform Network called
      NAP (Network Access Protocol) [4]. In discussions with NAP's
      originator, Gerry Bailey, the author was convinced of the
      fundamental reasonableness of the approach, but of course that
      doesn't have to convince others.  Additional rationales emerged in
      discussions with Gary Grossman, the originator of the DCA/DTI
      H-FP [5], which is probably the best-known current H-FP and which
      furnished the default Channel Layer for the new one:  In the first
      place, the text approach makes parsing for the ends of
      variable-length parameters easier.  In the second place, it allows
      for the possibility of creating a terminal-supporting OPE in a
      very straightforward fashion should any OPE developer elect to do
      so.  (See below for more on the distinction between OPE developers
      and H-FP implementors.) Finally, there's nothing sacred about
      binary headers anyway, and just because the text approach is
      different doesn't make it "wrong".  So, although it's not out of
      the question that the new protocol should back off from the text
      approach if reviewers and/or implementors come up with compelling
      reasons for doing so, the already frequently encountered reaction
      of "it feels funny" isn't compelling.  (It was, indeed, the
      author's own initial reaction.)  Besides, "nobody" (not even Gary)
      really liked the top layer of the DCA/DTI H-FP.

   X.25 Appropriateness

      Of more concern than how text "feels" is whether X.25 "works".
      That is, we understand that many system proprietors would greatly
      prefer being able to use "off-the-shelf" software and hardware to
      the greatest extent feasible and still be able to do intercomputer
      networking according to DOD Standards, which is a major reason why
      we decided to take the H-FP commands out of the Channel Layer of
      the DCA/DTI H-FP even before we decided to encode them as text.
      However, it is by no means clear that any old vendor supplied
      "X.25" will automatically be usable as a new H-FP Channel and Link
      layer mechanization.  As noted, it all depends upon how Host


Padlipsky                                                       [Page 7]


RFC 928                                                    December 1984
Introduction to H-FP


      programs (the Command Layer/H-FP Protocol Interpreter in
      particular) are able to invoke X.25 on particular systems.  Also,
      there might be peculiarities in the handling of some constructs
      (the Group and Member fields--or whatever they're called--are a
      strong candidate) which could militate against getting JUST
      demultiplexing and flow control out of X.25-as-Channel
      Link/Layers.  For that matter, it's conceivable that on some
      systems only one process can "own" the presumed DCE, but there's
      no interprocess communication available between it and the
      processes that want to use H-FP.  What that all amounts to, then,
      is that we don't pretend to be sufficiently versed in the vagaries
      of vendor-idiosyncratic X.25 implementations to claim more than
      that we THINK the new H-FP Command Layer should fit "on top of"
      X.25 in a Host such that a suitably crafted OPE could look like a
      DCE to the low-level Host software and still be an OPE in our
      sense of the term.  Finally, some reports on bit-transfer rates
      attainable through typical X.25 interfaces give rise to concern as
      to whether such a lash-up would be "good" even if it were
      feasible.

      DCA/DTI Channel Layer Appropriateness

      The Channel Layer of the DCA/DTI H-FP has been implemented for a
      few Host types already, and is being implemented for others (in
      particular, as part of the DODIIS NFE project). A delicate
      decision is whether to alter the header structure (e.g.--and
      perhaps i.e.--to remove the now-superfluous command and response
      fields).  On the "con" side are the considerations that
      implementations DO exist, and that it's well specified.  On the
      "pro" side are that keeping the header as it is is in some sense
      "wasteful" and that somebody's going to have to go over the spec
      again anyway, to remove that which no longer applies.  (It should
      be noted that Gary Grossman was initially tempted to scuttle the
      Group and Member trick, but the presence of a similar
      dichotomizing in X.25 seems to rule that out.)  One of the
      interesting issues during the review phase of the new H-FP, then,
      will be the decision about which way to go on the Channel Layer
      header in its non-X.25 version.  (NOBODY considers going X.25
      only, be it noted.)  By the time the protocol is finalized, it
      will, of course, be made clear in the protocol spec, but I'll
      probably leave this in the final version of the Introduction just
      for historical interest anyway.

   Syntax

      Another point which probably needs close scrutiny during the
      review process is the "syntax" of the command lines.  Basically,


Padlipsky                                                       [Page 8]


RFC 928                                                    December 1984
Introduction to H-FP


      we just took our best shot, but without any claims that it's the
      best possible way to express things.  So comments and/or
      alternatives are earnestly solicited on this one.

   L III Offloading

      Contrary to the expectations of some, we are allowing for the
      offloading of Process/Applications Layer (ARPANET Reference Model
      L III) protocols.  Both Bailey and Grossman reported favorably on
      the feasibility of this. Two points should be made, however: It's
      perfectly fair for a GIVEN OPE implementation not to offload a
      given L III protocol, although it would presumably not sell as
      well as ones which did.  That is, we're not claiming that by
      inventing a mechanization of the feature in the spec we levy a
      constraint on everybody who implements "the protocol", (Cf.
      Fabrication under Non-Issues, below). Just as we were feeling our
      way on syntax in general, we're really feeling our way when it
      comes to the L III stuff.  (I'm not even sure I managed to convey
      what I meant for "mediation level" to Joel and Dick.)  Again,
      suggestions are solicited.

   Security

      During the detailed design pass, we had an intensive discussion
      with some of the Blacker design team on the interplay between the
      new H-FP and a meant-to-be multilevel-secure OPE such as Blacker.
      The conclusion was that by and large "Security" is to be an aspect
      of an enhanced H-FP, rather than the standard one. The reasoning
      was rather involved, but seems to amount to the following:  Hosts
      that are NOT MLS (or "Compartmented") have two significant
      properties in our context: They're in the vast majority of
      present-day systems.  They have no legitimate need even to tell
      their OPEs what they "think" their current System High or
      Dedicated Mode level is; that information should be furnished by
      some trusted portion of a network security architecture (e.g., a
      security enhanced OPE, or a table in a "secure" comm subnet
      processor).

      Thus, even having the optional security label/level field in the
      Begin command is in some sense overkill, because we're not sure of
      any sensible circumstances in which it would be useful, but we put
      it in "just in case".  On the other hand, Hosts that ARE
      MLS/Compartmented by definition can be permitted to assert what
      the level of a given transmission (or perhaps of a given
      connection) should be, and their OPEs need to have a mechanism for
      learning this.  But it is by no means clear that a given Host (or
      even a given OPE) will be so structured as to make the H-FP PI,


Padlipsky                                                       [Page 9]


RFC 928                                                    December 1984
Introduction to H-FP


      the Channel PI, and the Link PI ALL trustworthy--as they'd have to
      be if the security labeling were part of H-FP.  So, we envision
      the labeling's being handled by trusted code in both Host and OPE
      that will be inserted into the normal processing route at the
      appropriate point for the given architecture (presumably "at the
      very bottom" of the Host, and "the very top" of the OPE), and that
      will place the label in a convenient, known position in the
      Host-OPE transmission "chunk" (block/packet/data unit) as the
      circumstances dictate. (It's likely--but we wouldn't swear to
      it--that a good place would be just before the H-FP command, and
      if that's the case then semi-clearly the security enhanced H-FP
      PIs would have to "make room" for it in the sense of handing the
      Channel Layer a suitably lengthened "chunk".)

      The Host and its OPE should be viewed as a single entity with
      regard to labeling requirements in the non-MLS/C case, and either
      the OPE will be conditioned to emit the right label or the CSNP
      will "know" anyway; in the MLS/C Host and OPE case (and it should
      be noted that it's just about impossible to envision a MLS/C Host
      which IS outboarded which DOESN'T have a MLS/C OPE) it will depend
      on the given security architectures as to whether each "chunk"
      needs labeling (i.e., there COULD be trusted H-FP, Channel, and
      Link PIs, so that only at channel establishment time does the
      label need to be passed), but it seems likely each "chunk" would
      need labeling, and we can see how that would happen (as sketched
      above).

      This is all, of course, subject to reappraisal when the full-time
      Security folks get in the act, but for now, H-FP per se is viewed
      as playing no direct role in "Security"--except indirectly, as
      noted below under the Symmetric Begins Non-Issue.  (In case
      anybody's worrying about the case where the OPE is physically
      remote from its Host, by the way, that line would have to be
      protected anyway, so the Host/OPE-asa-single-unit view should hold
      up.)

   How It Implements

      The final issue to take note of is that one of the central
      premises of the Outboard Processing approach has always been that
      H-FPs can be invented which implement more compactly on the Host
      side than the code they're allowing to be offloaded.  We certainly
      think the new H-FP will fulfill that condition, but we'd certainly
      like to hear of any evidence to the contrary.





Padlipsky                                                      [Page 10]


RFC 928                                                    December 1984
Introduction to H-FP


Non-Issues

   The following items are declared to be non-issues, in the sense that
   even though some people have expressed concern over them we believe
   that they are either "not part of the protocol" or resolved already
   for reasons that were overlooked by those worried about them:

   Fabrication

      Who builds OPEs isn't within our purview, except to the extent of
      hoping a few volunteers come forward to do testcase
      implementations of what is, at present, only a paper protocol.
      However, beyond agreeing that a few points should be marked as
      "Notes to Entrepreneurs" in the spec, we didn't attempt to dictate
      how an OPE vendor would behave, beyond the explicit and implicit
      dictates of the protocol per se. For example, if a given OPE
      doesn't offload SMTP, it jolly well ought to respond with the
      appropriate "Function not implemented" code, and if a vendor
      claims to accept X.25 for Channel and Link disagreements over what
      X.25 "is" are the province of the vendor and the customer, not of
      the H-FP spec.  As OPE'S are supposed to be offloading COMMON
      protocols in a COMMON fashion, a given OPE should be able to
      interoperate with another Host irrespective of whether that Host
      even has an OPE, much less whose OPE it is if it's there. Thus,
      for example, even though you'd expect to find OPEs that "come
      with" their own LANs as a fairly frequent product, we don't appeal
      to the notion in the conceptual model; nor do we attempt to
      dictate "chunk" sizes at the Channel level. A protocol spec isn't
      an implementation spec.

   Symmetric Begins

      For almost as long as there have been H-FPs, there has been
      disagreement over whether only the Host can begin a connection or
      if the OPE can also take the initiative.  I am delighted to be
      able to resolve this one finally:  It turns out there IS a
      compelling reason for insisting that THE PROTOCOL include
      provision for OPE --> Host Begins, so it's "in" the protocol--but
      any Host that doesn't need to deal with them doesn't have to (just
      "spell" the "Function not implemented" response code correctly).

      (In case anybody cares, the compelling reason is that if you HAD
      an MLS OPE which happened to use a security kernel and a process
      per level, you'd need IT to be listening for incoming connection
      requests "from the net" rather than having the Host tell it to do
      so, for various esoteric reasons--but in order to cater to the
      possibility, we want the function in the protocol from the


Padlipsky                                                      [Page 11]


RFC 928                                                    December 1984
Introduction to H-FP


      beginning, on the grounds that we can envision SOME other uses for
      it even in non-MLS environments [unlike the security labeling
      trick discussed above, which only seems to make sense for MLS
      Hosts/OPEs--that is, it doesn't burden the Host to reject a Begin
      every once in a while but it would to go around labeling "chunks"
      unnecessarily all the time].)

   Routing

      Concern has been voiced over the issue of what provisions the
      protocol should make to deal with the situation where a Host,
      probably for traffic/load reasons, has multiple OPEs and the
      question arises of which OPE to use/route to.  I claim this is a
      non-issue at the protocol level.  If the Host-side H-FP PI gets a
      "No resources" response to a Begin, it can go off to another OPE
      if it wants to.  "Not our department".  The conceptual model is
      that of a Host and AN OPE--which "ought to" be expandable to carry
      more load at some level.  If you want multiple links for some
      reason, the simplest solution would seem to be to have multiple
      Channel Layers as well, but the whole thing just gets too iffy to
      have anything sensible to prescribe in the protocol.  In other
      words, extending the concept to deal with discrete multiple OPEs
      is either a Fabrication sort of thing, or a Notes to Host-side
      Implementors sort of thing on a per specific OPE basis.

   Operator Interface

      It's probably implicit in the foregoing, but it might be worth
      saying explicitly that the operator interface to a specific OPE is
      a non-issue in terms of the protocol, beyond the provision we're
      made for "Shutdown coming" responses as a reflection of a probable
      operator interface action we imagine most operator interfaces
      would provide.  (It might also be worth noting that if your Host
      does "color changes", your OPE had better have a trustworthy way
      of being told to change the label it plops on all IP datagrams it
      emits, but that comes under the heading of an Aside to Specialized
      Implementors.)












Padlipsky                                                      [Page 12]


RFC 928                                                    December 1984
Introduction to H-FP


Fine Points

   There are a couple of known "loose ends" which are exceedingly fine
   points in some sense that do bear separate mention:

   The Allocate Event

      While mentally testing to see if the new H-FP would indeed
      off-load TCP, we came up against an interesting question: Viewing
      H-FP as "just an interface at a distance" to a TCP PI, what about
      the Allocate "Interface Event" in the TCP spec?  As far as I'm
      concerned, this could be classed as a non-issue, because I submit
      that the spec is wrong in declaring that there is such a thing as
      a MANDATORY Interface Event whereby the user of a TCP PI lets the
      PI know how much data it can take. Granted, you might find such a
      thing in most implementations, but what if you were in a virtual
      memory environment with segment sharing (or a distributed
      supervisor) and you wanted to avoid copies, so all that passed at
      the interface to the PI (or even at the interface from the PI) was
      a pointer?  That is, the "DOD version" of the TCP spec has fallen
      into the trap of assuming things about the execution environment
      that it shouldn't have.

      One moral of this is that

         AN INTERFACE TO AN INTERPRETER OF A PROTOCOL IS N*O*T "THE
         PROTOCOL".

      Another moral is that the interface to the Host-side H-FP PI is
      hard to say much about, but is where the equivalent functionality
      will be found if you've offloaded TCP.  That is, it's reasonable
      to let the user "tell" the outboard PI at Begin time if big or
      small buffers are expected to be in play "net-ward" as part of the
      protocol, but the outboard PI is expected to deliver bits to the
      Host as they come unless throttled by the Channel Layer, or by
      some to-be-invented other discipline to force the OPE to buffer.
      (For present purposes, we envision letting the Channel Layer
      handle it, but nifty mechanizations of encouraging the OPE to
      "make like a buffer" would be at least looked at.)  As a
      Fabrication issue, it is the case that "equity" has to be dealt
      with with regard to the use of the OPE's resources (especially
      buffers) across H-FP connections/channels, but that's a different
      issue anyway, touched upon in the final fine point.






Padlipsky                                                      [Page 13]


RFC 928                                                    December 1984
Introduction to H-FP


   Precedence

      Clearly, the existence of a notion of Precedence in DOD protocols
      has to get reflected in the outboard PI's implementations. Just
      what, if any, role it has in the H-FP, per se, is, however, by no
      means clear.  That is, if the Host doesn't take Begins from the
      OPE and is "full up" on the number of Server Telnet connections
      it's willing to handle, what should happen if a high precedence
      SYN comes in on the Telnet Well-Known Socket (in present day
      terms)?  Probably the OPE should arbitrarily close a low
      precedence connection to make room for the new one, and signal the
      Host, but even that assumes the Host will always hurry to be
      prepared to do a new passive Begin.  Perhaps we've stumbled across
      still another argument in favor of "Symmetric Begins"....  At any
      rate, Precedence does need further study--although it shouldn't
      deter us from making "the rest" of the protocol work while we're
      waiting for inspiration on how to handle Precedence too.

A Note on Host Integration

   The most important thing about Hosts in any intercomputer network is
   that they furnish the resources to be shared. The most significant
   obstacle to sharing those resources, however, is the fact that almost
   invariably they were designed under the assumption that the Host was
   a fully autonomous entity.  That is, few operating systems currently
   deployed "expect" to be members of a heterogeneous community of
   operating systems.  In many cases, this built-in insularity goes so
   far as to have applications programs cognizant of the particular type
   of terminal from which they will be invoked.

   Intercomputer networking protocols attempt to resolve the problems of
   heterogeneity by virtue of presenting appropriate common intermediate
   representations (or "virtualizations") of the constructs and concepts
   necessary to do resource sharing.  A Host-Host protocol such as TCP
   "is" a virtual interprocess communication mechanism; a virtual
   terminal protocol such as Telnet obviously is a mechanism for
   defining and dealing with virtual terminals; FTP offers common
   representations of files; and so on.  It cannot be stressed strongly
   enough, though, that this entire approach to intercomputer networking
   is predicated on the assumption that the modules which interpret the
   protocols (PIs, as we'll refer to them often) will be PROPERLY
   integrated into the various participating operating systems.  Even in
   the presence of powerful OPEs, wherein the bulk of the work of the
   various PIs is performed outboard of the Host, the inboard "hooks"
   which serve to interface the outboard PIs to the native system must
   not only be present, they must be "right".  The argument parallels
   the analysis of the flexible vs. rigid front-ending attachment


Padlipsky                                                      [Page 14]


RFC 928                                                    December 1984
Introduction to H-FP


   strategy issue of [1]; to borrow an example, if you attempt to
   integrate FTP by "looking like" a native terminal user and the
   operator forces a message to all terminals, you've got an undetected
   pollution of your data stream. So the key issue in attaching Hosts to
   networks is not what sort of hardware is required or what sort of
   protocol is interpreted by the Host and the OPE (or comm subnet
   processor, for that matter), but how the PIs (full or partial) are
   made to interrelate with the pre-existing environment.

   It would be well beyond the scope of this document to attempt even to
   sketch (much less specify) how to integrate H-FP PIs into each type
   of operating system which will be found in the DoD.  An example,
   though, should be of use and interest.  Therefore, because it is the
   implementation with which we are most intimately familiar, even
   though it's been several years, we propose to sketch the Multics
   operating system integration of the original ARPANET Network Control
   Program (NCP)--which is functionally equivalent to an H-FP PI for
   offloading ARM L II and L I--and Telnet.  (A few comments will also
   be made about FTP.) Note, by the way, that the sketch is for a
   "full-blown" H-FP; that is, shortcuts along the lines of the
   scenario-driven approach mentioned above are not dealt with here.

   One of the particularly interesting features of Multics is the fact
   that each process possesses an extremely large "segmented virtual
   memory".  That is, memory references other than to the segment at
   hand (which can itself be up to 256K 36-bit words long) indirect
   through a descriptor segment, which is in principle "just another
   segment", by segment number and offset within the segment, so that a
   single process--or "scheduling and access control entity"--can
   contain rather impressive amounts of code and data.  Given that the
   code is "pure procedure" (or "re-entrant"), a "distributed
   supervisor" approach is natural; each process, then, appears to have
   in its address space a copy of each procedure segment (with
   system-wide and process-specific data segments handled
   appropriately).  Without going too far afield, the distributed
   supervisor approach allows interrupts to be processed by whichever
   process happens to be running at a given time, although, of course,
   interprocess communication may well be a consequence of processing a
   particular interrupt.

   A few other necessary background points:  A distinguished process,
   called the Answering Service, exists, originally to field interrupts
   from terminals and in general to create processes after
   authenticating them.  Other shared resources such as line printers
   are also managed by distinguished processes, generically known as
   "Daemons".  Device driver code, as is customary on many operating
   systems, resides at least in part in the supervisor (or hard core


Padlipsky                                                      [Page 15]


RFC 928                                                    December 1984
Introduction to H-FP


   operating system).  Finally (for our purposes, at least), within a
   process all interfaces are by closed subroutine calls and all I/O is
   done by generic function calls on symbolically named streams; also,
   all system commands (and, of course, user written programs which need
   to) use the streams "user_input" and "user_output" for the obvious
   purposes.  (At normal process creation time, both user I/O streams
   are "attached" to the user's terminal, but either or both can be
   attached to any other I/O system interface module instead--including
   to one which reads and writes files, which is handy for consoleless
   processes.)

   All that almost assuredly doesn't do justice to Multics, but equally
   likely is more than most readers of this document want to know, so
   let's hope it's enough to make the following integration sketch
   comprehensible. (There will be some conscious omissions in the
   sketch, and doubtless some unconscious ones, but if memory serves, no
   known lies have been included.)

   Recalling that NCP is functionally equivalent to H-FP, let's start
   with it. In the first place, the device driver for the 1822 spec
   hardware interface resides in the supervisor. (For most systems, the
   PI for H-FP's link protocol probably would too.)  In Multics,
   interrupt time processing can only be performed by supervisor
   segments, so in the interests of efficiency, both the IMP-Host (1822
   software) Protocol PI and the multiplexing/demultiplexing aspects of
   the Host-Host Protocol PI also reside in the supervisor.  (An H-FP PI
   would probably also have its multiplexing/demultiplexing there; that
   is, that portion of the Channel Layer code which mediates access to
   the OPE and/or decides what process a given message is to be sent to
   might well be in the supervisor for efficiency reasons.  It is not,
   however, a hard and fast rule that it would be so. The system's
   native interprocess communications mechanism's characteristics might
   allow all the Channel Layer to reside outside of the supervisor.)

   Even with a very large virtual memory, though, there are
   administrative biases against putting too much in the supervisor, so
   "everything else" lives outside the supervisor. In fact, there are
   two places where the rest of the Host-Host Protocol is interpreted on
   Multics, although it is not necessarily the case that an H-FP PI
   would follow the same partitioning even on Multics, much less on some
   other operating system.  However, with NCP, because there is a
   distinguished "control link" over which Host-Host commands are sent
   in the NCP's Host-Host protocol, the Multics IMP-Host Protocol PI
   relegates such traffic to a Network Daemon process, which naturally
   is a key element in the architecture.  (Things would be more
   efficient, though, if there weren't a separate Daemon, because other
   processes then have to get involved with interprocess communication


Padlipsky                                                      [Page 16]


RFC 928                                                    December 1984
Introduction to H-FP


   to it; H-FP PI designers take note.)  To avoid traversing the Daemon
   for all traffic, though, normal reads and writes (i.e., noncontrol
   link traffic) are done by the appropriate user process.  By virtue of
   the distributed supervisor approach, then, there is a supervisor call
   interface to "the NCP" available to procedures (programs) within user
   processes. (The Daemon process uses the same interface, but by virtue
   of its ID has the ability to exercise certain privileged primitives
   as well.)

   If a native process (perhaps one meaning to do "User Telnet", but not
   limited to that) wanted to use the network, it would call the open
   primitive of "the NCP", do reads and writes, and so on.  An
   interesting point has to do with just how this interface works:  The
   reads are inherently asynchronous; that is, you don't know just when
   the data from the net are going to be available.  In Multics, there's
   an "event" mechanism that's used in the NCP interface that allows the
   calling process to decide whether or not it will go blocked waiting
   for input when it reads the net (it might want to stay active in
   order to keep outputting, but need to be prepared for input as well),
   so asynchrony can be dealt with.  In the version of Unix (tm) on
   which an early NFE was based, however, native I/O was always
   synchronous; so in order to deal with both input from the terminal
   and input from the net, that system's User Telnet had to consist of
   two processes (which is not very efficient of system resources).
   Similar considerations might apply to other operating systems
   integrating H-FP; native I/O and interprocess communication
   disciplines have to be taken into account in designing.  (Nor can one
   simply posit a brand new approach for "the network", because Telnet
   will prove to rely even more heavily on native mode assumptions.)

   The other aspect of NCP integration which we should at least touch
   on--especially because process-level protocols make no sense without
   it--is how "Well-Known Sockets" (WKSs) work. In broad terms, on
   Multics the Network Daemon initially "owns" all sockets.  For
   Well-Known Sockets, where a particular process-level protocol will be
   in effect after a successful connection to a given WKS, code is added
   to the Answering Service to call upon the NCP at system
   initialization time to be the process "listening" on the WKSs.  (This
   is a consequence of the fact that the Answering Service is/was the
   only Multics process which can create processes; strategies on other
   systems would differ according to their native process creation
   disciplines.)  How to get the "right kind of process" will be
   sketched in the discussions of the process level protocols, but the
   significant notion for now is that typically SOME sort of prior
   arrangement would be done by any networked Host to associate the
   right kind of process with a WKS.



Padlipsky                                                      [Page 17]


RFC 928                                                    December 1984
Introduction to H-FP


   Now, we don't expect that the foregoing will enable even the world's
   greatest system jock to go out and design the integration of an H-FP
   PI for a system that had never been networked (in the ARPANET style
   of networking) before. But we propose to stop there and turn to some
   comments on process level protocols, for two reasons: In the first
   place, it would take us much too far afield to go into significantly
   greater detail; and in the second place, because of the functional
   equivalence of H-FP and NCP combined with the number of operating
   systems which have integrated NCP and, for that matter, TCP/IP, which
   are also functionally equivalent to H-FP (used for offloading L II
   and L I), models are available in the ARPANET community and concerned
   H-FP PI implementors can follow them.

   Turning to Telnet integration, and returning to Multics as an
   example, we note that "User Telnet" is straightforward. "All you
   need" (for small values of "all") from an INBOARD User Telnet is a
   command that gives the user some sort of interface, converts between
   the native Multics character set and terminal discipline and the
   Network Virtual Terminal equivalents (and as Multics is very generic
   when it comes to I/O, that's not hard), and writes and reads "the
   net" (more accurately, calls upon the Host-Host protocol PI--or upon
   the H-FP PI to get at the H-HP--appropriately).  (One point that's
   not obvious:  make the Well-Known Socket "on the other side" a
   parameter, defaulting to the Telnet WKS, because you'll want to use
   the same command to get at other process-level protocols.)  If
   there's an OPE in play which offloads User Telnet, however, things
   can be even simpler: the inboard command just reads and writes the
   terminal and lets the OUTBOARD User Telnet PI handle the conversion
   to and from the Virtual Terminal form (presumably, from and to the
   desired local form).

   When it comes to the incoming ("Server") aspects of Telnet, life can
   get complicated on some systems for an inboard implementation.
   However, fortunately for our purposes,

   Multics' native mechanisms lend themselves readily to integration; an
   awareness of the inboard issues will be useful even if in response to
   a connection attempt on the Telnet WKS, the (Server) Host is
   obligated to associate the connection (the actual logic is somewhat
   more complex under the ARPANET Host-Host Protocol, which employs
   paired simplex connections) with a process that is prepared to
   translate between Telnet and native mode representations and
   otherwise "look like" a local user process--that is, in particular
   the connection becomes an I/O source/sink to the native command
   processor on time-sharing systems.  As indicated, process creation is
   taken care of in Multics by having the Answering Service process
   listen on the WKS.  Because the Answering Service is in some sense


Padlipsky                                                      [Page 18]


RFC 928                                                    December 1984
Introduction to H-FP


   just another Multics process, it too does user I/O through the normal
   system mechanisms.  So while for local terminals the user I/O streams
   are attached through a module called "ttydim" (where "dim" stands for
   "device interface module"), NVTs are attached through a functionally
   equivalent and identically invoked module called "nttydim" (the
   Answering Service knows which DIM to use based on the symbolic
   designator of the "line" on which it received the interrupt, as it
   happens).

   [The notion of "attaching" the streams bears a bit more explanation:
   Attach is a primitive of the Multics generic I/O mechanism which
   associates a stream name and a particular DIM (or I/O system
   interface module in later terminology); the other I/O primitives
   (read, write, etc.) are invoked with the stream name as a parameter
   and an I/O "switch" causes the entry point corresponding to the
   primitive to be invoked in whichever DIM the stream is currently
   attached to.   So a Server Telnet process starts life attached
   through nttydim to a particular network connection, while a local
   process starts life attached through ttydim to a particular physical
   line, and both processes proceed indistinguishably (viewed from
   outside the I/O switch, anyway).]

   The pre-existing orderliness that makes things easy on Multics does
   not, unfortunately, appear in all operating systems.  Indeed,
   delicate choices occasionally have to be made as to WHICH native
   terminal to map to on systems that don't do generic I/O in native
   mode, and it is likely that for some systems the particular mapping
   to bring into play in Server Telnet might be determined by the
   particular application program invoked.  This issue can become very
   touchy when the application "expects" a "data entry terminal", say.
   The Server Telnet for such a system would naturally attempt to
   negotiate the "DET" option with the corresponding User Telnet.  But
   the user might be at a physical terminal that isn't a member of the
   DET class, so that User Telnet must either refuse to negotiate the
   option or--and we would recommend this alternative strongly, as it
   seems to be within the "spirit" of the protocol--offer some sort of
   simulation, however crude, of the behavior of a DET.   Also,
   something sensible has to be done on systems where there is no clear
   analog of the command processor expected to be managing the Server
   process.  (Say, when a "menu" of applications is always displayed on
   an available terminal in native mode.)

   A final Telnet integration issue (although other points could be
   noted, we're not pretending to be exhaustive and this should be
   enough to "give the flavor"):  The Telnet Interrupt Process generic
   function calls for particularly careful integration.  Here, the
   intent of the function is to virtualize what is called the "quit


Padlipsky                                                      [Page 19]


RFC 928                                                    December 1984
Introduction to H-FP


   button" on some systems. That is, the user wants the system to
   interrupt his process (which may, for example, be in a loop) and get
   back to the command processor (or "the system" itself).   On native
   character-at-a-time systems, the native mechanism is usually the
   entering of a particular "control character"; on native
   line-at-a-time systems, the native mechanism is usually the striking
   of the "ATTN" or Interrupt button or the "Break" key (sometimes more
   than once, to distinguish it from a communication to the executing
   program).  But the native mechanisms typically involve interrupt time
   code, and Server Telnet typically wouldn't be executing at that
   level, so the solution (omitting the intricacies of the interaction
   with the NCP or the H-FP PI, which also get into the act) would be to
   make use of--in the Multics case--a pre-existing INTRAprocess signal,
   or to add such a mechanism (unless the architecture chosen has a
   Server Telnet Daemon of some sort, in which case an INTERprocess
   signal would be needed).

   The extension of the foregoing to an outboard Server Telnet may not
   be obvious, but we won't expend a great deal of time on it here.
   Even if "the protocol" is being handled in an OPE, the Host-side
   software must be able to associate an H-FP connection with the
   command language interpreter of a user process and to respond
   appropriately to an H-FP Signal command if it arrives, and the OPE
   must know not only the desired character set but also the local
   equivalents of Erase and Kill, at the minimum.

   We'll skip FTP integration, on the grounds that this note is already
   too lengthy, except to mention that in the OUTBOARD case it's still
   going to be necessary to convey the name of the appropriate file and
   directory to/from some appropriate Host-side code.  (Similar problems
   must be dealt with for outboard handling of "mail" if it's not part
   of FTP.)

   One other "integration" issue, which has been hinted at earlier and
   about which not much can be said beyond some general guidelines: The
   "top edge" of a Host-side H-FP protocol interpreter (i.e., the Host
   user program interface, for

   Hosts that are "doing real networking" rather than just using the OPE
   to get at User Telnet and/or FTP and to offer Server Telnet and/or
   FTP [and maybe "mail"], presumably in the "scenario-driven" fashion
   sketched earlier) MUST BE APPROPRIATE TO THE HOST.  In other words,
   on Multics, where "everything" is closed subroutines, there would
   presumably be a closed subroutine interface with event channels for
   reads, pointers to buffers, and all that sort of thing, but on some
   other style of operating system, the interface to the H-FP PI might
   turn out to be "all" interprocess communication, or to "look like" a


Padlipsky                                                      [Page 20]


RFC 928                                                    December 1984
Introduction to H-FP


   device of some special class, or "all" system
   calls/JSYSs/EOTs/Whatevers.  We can't be much more specific, but we'd
   be remiss to convey any impression that H-FP is a "free lunch".  As
   noted, an H-FP PI requires the same kind of integration as a generic
   NCP--it's just smaller, and serves as insulation against changes (in
   the offloaded protocols in general, or in the proximate comm subnet
   in particular).

References

   (References [1]-[3] will be available in M. A. Padlipsky's "The
   Elements of Networking Style", Prentice Hall, 1985.)

   [1] Padlipsky, M. A., "The Host-Front End Protocol Approach", MTR
   3996, Vol. III, MITRE Corp., 1980.

   [2] Padlipsky, M. A., "The Elements of Networking Style", M81-41,
   MITRE Corp., 1981.

   [3] Padlipsky, M. A., "A Perspective on the ARPANET Reference Model",
   M82-47, MITRE Corp., 1982.

   [4] Bailey, G., "Network Access Protocol", S-216,718, National
   Security Agency Central Security Service, 1982.

   [5] Day, J. D., G. R. Grossman, and R. H. Howe, "WWMCCS Host to Front
   End Protocol", 78012.C-INFE.14, Digital Technology Incorporated,
   1979.





















Padlipsky                                                      [Page 21]