GENWiki

                                                             J. Postel

IEN 145 ISI

                                                           29 May 1980

             Internet Meeting Notes - 14 & 15 May 1980

I. INTRODUCTION

 The meeting was held at the Laboratory for Computer Science (LCS) at
 MIT.  Dave Clark was the host.

II. OVERVIEW AND OBJECTIVES

 Vint Cerf put forward the problem of internet computer mail as a
 topic that must be addressed in the next few months.  This would be
 an interim system not expected to support multimedia mail but to
 operate on TCP and to provide some mechanism for interworking NCP
 supported mail with TCP supported mail.

 Another issue is gateway-host interaction, especially for congestion
 control.

 A third topic is routing in a large packet radio environment with
 optional use of backbone point-to-point links.

III. STATUS REPORTS

 A.  ARPA

    Vint Cerf reported that the IP and TCP specifications (IENs 128
    and 129) are now stamped as DOD standards.  DCEC will serve as the
    focus for coordinating this within DOD.  A seminar for DOD people
    is planned for July at NBS to explain TCP.

    The monthly reports are very useful, but still needed are some
    milestone schedules, for planning and cross project coordination.

 B.  BBN

    Dale McNeill reviewed the work on the VAN gateway, noted the
    arrival of the FAX equipment, described the installation of the
    PSP terminal at TANUM, and noted that improvements have been made
    to the CMCC software.

    Mike Brescia described the VAN gateway in more detail.  The
    operation of the gateway is still some way off.  Current activity

Postel [Page 1]

                                                               IEN 145

Internet Meeting Notes

    focuses on the X.25 interfaces.  At this point there was an
    extended discussion of low level interface options.

    Jack Haverty described the work on the Unix IP and TCP.  Most
    recently a version using shared memory for user/TCP communication
    has been tested and is 2 to 3 times faster than the older "pipes"
    version.  Work is just beginning on an IP and TCP for VAX Unix.
    This version will be in the kernel and will include the
    fragmentation and reassembly features.  Also in progress are a TCP
    for a HP 3000, and a TCP-TIP in a 316.  There was also some
    discussion of which versions of Unix are involved:  The VAX Unix
    is version 7, the current 11/70 Unix is version 6.  Jack also
    mentioned some performance measurements using a traffic generator.
    This will be documented in a forthcoming IEN.

    Ruth Nelson briefly discussed a local net project at BBN which is
    building a net based on the CHAOS net design.  There is now a
    system with two interfaces which is tested by linking two
    terminals.

    Bill Plummer reported on the TENEX/TOPS20 IP and TCP status.  A
    number of improvements have been made to the debugging version at
    BBNF and these should soon be distributed and installed at other
    sites.  Among the improvements are:  a resolution of the "data
    stream capture problem," an update to the internet user queues, an
    improvement to the gateway functions - especially updating the
    gateway table.  Bill has also prepared an "Installation Guide for
    Wizards."

    Ginny Strazisar reported that the last ELF based gateway has been
    converted to MOS, and that the first gateway running in an LSI-11
    has been delivered to SRI.  Also a three port gateway is working.

 C.  COMSAT

    Hoi Chong reported that there are currently a few problems with
    the COMSAT gateway which seem to be related to a power supply.
    Also it was noted that the line to NSRDC will be removed at the
    end of May.  The IP and TCP used in the COMSAT hosts has been
    improved and these programs now support the timestamp option.  The
    FAX machine and protocols are in place and ready to be used, what
    is needed is a common FTP to move the bits to other sites.  COMSAT
    was particularly active in the "Bakeoff" as reported in several
    messages.  Within COMSAT there is an investigation of a local net
    to link some PDP-11 and IBM computers.  Hoi also distributed a
    memo on COMSATs internet activities and milestones.

Postel [Page 2]

                                                               IEN 145

Internet Meeting Notes

 D.  DCA

    Ed Cain reported that the DCEC gateway had been considerably
    improved with help from Bill Plummer, Dave Mills and others.  Ed
    also reported that DCEC role in the DOD protocol standards will
    have three aspects:  (1) Executive Agent, (2) Review Panel, (3)
    Laboratory.  This Laboratory will be based on the EDN.

 E.  DOD

    Ray McFarland reported that a group in his office building a TCP
    2.5 for a PDP 11/34 with Unix (v.6).  This group needs software
    for Telnet and Echo processes.  Ray also discussed his role as an
    information coordinator in DOD.  He receives many requests for
    information from within DOD.  He needs help in fielding these
    questions.  In particular more data on performance would be very
    helpful.

 F.  ISI

    Jon Postel reported that multi-media mail system is being
    redesigned, and that a new draft of the MPM specification is
    available.  Also good progress has been made in the investigation
    of program verification tools for protocol analysis - especially
    AFFIRM.  ISI has ordered 4 computer readable clocks which maintain
    their time by listening to WWVB.  Jon also described the recent
    movement of users between the various machines at ISI.

 G.  Linkabit

    Estil Hoversten reported that Linkabit is not much involved in the
    internet activities directly, but rather in the details of SATNET
    and the WBC project.  Estil and Danny Cohen are working on an
    overview paper.

 H.  LL

    Jim Forgie noted that one goal of his project is to demonstrate
    point-to-point speech via the WBCNET in FY 80.  This may be hard
    to do due to the slippage in the schedule for some of the key
    pieces.  Lincoln is going ahead with the development of the voice
    terminals and their access path -- the LEXNET.  In cooperation
    with ISI programs are being written to support NVP2, ST, and IP.
    Lincoln is using the ISI developed EPOS system for this project.

Postel [Page 3]

                                                               IEN 145

Internet Meeting Notes

 I.  MIT

    Dave Clark reported on the various systems at MIT and their
    interconnections.  Things are getting quite complex.  (1) Unix:
    MIT has a copy of the IP/TCP from BBN, which MIT modified to have
    a user accessible IP interface.  A TFTP was installed in this
    host.  (2) Multics has had only small changes since last time,
    there is now a Name Server.  (3) The ALTOs now have IP, UDP and
    TFTP in BCPL running; IP, TCP and Telnet (with SUPDUP) in MESA are
    being programmed.

    Dave noted that the lack of implementation of fragmentation and
    reassembly is causing problems.

    Dave gave some performance measures of the local network at MIT
    (LCSNET v1) which indicate very high reliability and very low
    average load.  However there is some trouble interfacing MIT-XX (a
    TOPS20) to this local net.

    Dave noted that version 2 of this network is in hardware testing
    and that the Nu Machine (a personal computer being developed at
    MIT) will interface to the version 2 LCSNET.

 J.  MITRE

    Anita Skelton reported that starting with a half completed Z8000 C
    cross-compiler obtained from MIT, the compiler was finished, an
    assembler and loader was written.  Then starting with BBN's C
    version of SRI's MOS, the interprocess communication was modified,
    the terminal handler was rewritten, and device drivers were added;
    MOS is running in the Z8000.  MITRE is now converting the MOS-TCP
    to C, rewriting large portions and following the DOD spec.

    There are two Z8000 development boards interfaced to the MITRE
    cable, with the cable contention algorithms coded, and packets
    have been exchanged on the cable.  MITRE hopes to have TCP running
    in the Z8000s soon.

    The next step is to implement Telnet and interface the 11/70 to
    the Z8000.  The intention is to have a high speed DMA interface to
    the 11/70 via the UMC-Z80 and also a parallel interface via the
    DR11-C, so that comparable measurements with TCP running over the
    cable can be made.  (These measurements for the older interface
    unit have been made).

    In addition, a cable bus test bed will be installed at DCA in
    Reston.  The interface units will be Z8000 based and built by
    Reaction Instruments; the interface units will sell for about

Postel [Page 4]

                                                               IEN 145

Internet Meeting Notes

    $4000 with a parallel host port and two terminal ports
    (asynchronous and synchronous).  The cable will sit between the
    IMP and the NFE, with terminals attached to the cable, and one
    Z8000 unit will act as a gateway.

 K.  NDRE

    Yngvar Lund reported that NDRE is investigating a local network
    with HDLC interfaces to the connected computers.  The connected
    computers will be built up of modules based on Z80 processors and
    programmed to support voice protocol, HDLC, TCP, etc.

 L.  RSRE

    Andy Bates noted the work on Extended Memory MOS was reported in
    IEN 136.  IP was recoded in CORAL66 for EMOS and a recoding of TCP
    is planned.  Modems to upgrade the RSRE-UCL line are on order.
    Work is in progress on a gateway between PPSN and PSS.

 M.  SRI

    Ron Kunzelman gave an overview of SRI's activities and a brief
    rundown on hardware available.  Ron discussed some instrumentation
    efforts underway at SRI.  A PDP 11/44 is on order to aid in this
    work.

    Ed Perry discussed some of the problems with the Ft. Bragg
    installation of the Packet Radio net.  The gateway (or attached
    nets) seem to have a low packet/second throughput limit.  This is
    causing the higher level protocols to be modified to use fewer
    packets (e.g., a "half duplex" user mode).

    There was some discussion at this point of XNET and again the
    problem of "big" datagrams.  It really is important to have
    fragmentation and reassembly implemented everywhere.

    Holly Nelson reported on the port expander.  New versions were
    delivered to Ft. Bragg and UCL.

    A new directory has been set up for the current version of TIU
    software.  It is <TIU-SOFTWARE> on SRI-KL.  News of interest to
    TIU programmer or others concerned with MOS software should be
    sent to the MOS-Users-List run by Noel Chiappa.

 N.  UCL

    Rob Cole related the recent activities at University College.
    There are many pieces of equipment at UCL connected in complicated

Postel [Page 5]

                                                               IEN 145

Internet Meeting Notes

    ways.  The current focus is a local "Cambridge Ring" network.  The
    NIFTP now works on ISIE and an LSI-11 host.  There is a revised
    specification of NIFTP forthcoming.  UCL has a grant to provide a
    computer mail service between UK and US university users.  There
    is a major concern about cost of transmission via IPSS, so current
    focus is on multi-destination mail in single transmission.

    The PDP-9s which serve as a gateway between the ARPANET and EPSS
    will die if either EPSS changes or the 32-bit leaders in the
    ARPANET go away.  The PIXIE device is no longer is use.

 O.  UCLA

    Bob Braden reported on the IP and TCP in the IBM 3033 at UCLA.
    The operating system is VM with OS-MVT.  The IP and TCP are
    running and the 96-bit leader (24 bit address) support is
    installed.  Further work is needed in the area of "gateway"
    functions in the IP.  In TCP some work is needed in the ACK policy
    area.  One can poke this system via telnet.  Try "NETSTAT" and
    "HELP-TCP".

 P.  XEROX

    Vint Cerf read the following report sent in by John Shoch:

    1.  We have contributed our efforts to measure the effectiveness
    of the Packet Radio Network when used as part of the Pup
    environment.  IEN 138 describes some of the (disappointing)
    results obtained with the new IPRs; we are in the process of
    moving one of the radios, however, and plan to replicate the tests
    in the near future.

    2.  IEN 140 is a paper prepared with Danny Cohen and Ed Taft,
    outlining the use of "mutual encapsulation" as a means to support
    the coexistence of the Pup protocols and the IP protocols.

    3.  On Tuesday, May 13, Xerox joined with DEC and Intel to
    announce the results of an effort to establish a common
    specification for a local network, based upon the Ethernet
    technology.

Postel [Page 6]

                                                               IEN 145

Internet Meeting Notes

IV. MEASUREMENTS

 Andy Bates reported that RSRE conducted further measurements (since
 the last meeting).  Their findings again indicate a large spread in
 the delay through the concatinated ARPANET and SATNET.  Some
 speculated that the ARPANET introduces this spread.  The RSRE group
 suggest that it would be appropriate to introduce a provision for a
 negotiated retransmission option.  Also a larger window would help.

 Rob Cole presented the results of some recent measurements made by
 UCL of the gateway and SATNET.  The gateway at UCL was able to echo
 packets at 34 packets/second.  When the path was extended to the BBN
 gateway, the UCL gateway accepted 14 packets/second, but very few
 echo replies were received from the BBN gateway.  A port expander in
 the path did not pose a limit on either measurement.

V. GATEWAY PROTOCOLS AND HOSTS

 Jim Mathis presented his procedure for routing.  The main points are
 first pick any gateway, second refine the chance to the best gateway
 and third detect the failure of that gateway should it occur.

    o  Pick a Prime gateway
    o  Poll it at a slow rate
    o  Send to the Prime gateway
    o  Accept and act on a Redirect message
    o  Ping gateway in use if higher level protocol complains
    o  Periodically change the Prime gateway

 Does this procedure get unstable in high load?

 IENs 109 and 131 should be reviewed by host IP implementors.

VI. TELNET AND FTP FOR TCP

 Jon Postel reported on draft specifications for Telnet and FTP for
 TCP.  For Telnet the key changes are that ICP is eliminated and the
 single full duplex connection is between ports U and L.  Another
 change is that the Telnet SYNCH becomes DM + Urgent, but one can't
 count Urgents in the way one used to count Interrupts.

 For FTP, again ICP is eliminated for the control connection, which is
 now the full duplex connection between ports U and L.  The major
 change here is the elimination of the BYTE command and all that
 implies.  Other changes affect the defaults for the data connection
 and the "third party" transfers.  The MAIL and MLFL commands are
 included in the specification, and the new reply codes are used.

Postel [Page 7]

                                                               IEN 145

Internet Meeting Notes

VII. NEW IMPS

 Vint Cerf reported that BBN has formed a separate BBN Computer
 Company.  This company sells a C-30 machine (formerly known as an
 MBB).  The current C-30 emulates an H316.

 Some things on the queue for IMP improvements are:

 1.  extend the memory capacity of the C-30 IMP program
 2.  make software changes to support more than 4 hosts
 3.  provide HDLC interfaces for hosts
 4.  investigate the non blocking interface
 5.  provide logical addressing for hosts

 Note also that a TCP-TIP is being developed with the TCP/Telnet code
 in a H316 and the IMP part in a C-30.

VIII. THE CMCC

 David Flood Page gave a brief review and a demonstration of the CMCC
 functions.  Basically the CMCC programs collect data from cooperating
 gateways and display the results on a terminal.  The programs reside
 at ISIE  in directory <CMCC>.  The key program is CMCC-DISPLAY.  Some
 files in this directory give helpful information about the program.
 These are:  HELP.TXT, NEWS.TXT, and SESSION.TXT.  Further information
 can be obtained from David via sndmsg to DFloodPage@BBNE.  Please see
 also IEN 131 and 132.  The demo went very well and these programs
 should be useful to any gateway builder.

IX. INTERNET PROTOCOL SPECIFICATIONS

 Ken Shotting presented some results of studying the IP specification
 for a formal description.  The key result is the identification of
 ambiguous areas in the specification.  In particular, the interaction
 of the fragmentation procedure and the return route option is one
 cause for concern.  Another issue is the use of the identification
 field.

X. EGG BREAKING

 Danny Cohen led a discussion of the problems arising from assumptions
 about which end of a word/page/..., bytes are transmitted from.  This
 is a holy war between the big-endians and the little-endians (see IEN
 137).  Noel Chiappa and Danny were appointed as a small group to
 argue about it.

Postel [Page 8]

                                                               IEN 145

Internet Meeting Notes

XI. FLYING PACKET RADIOS

 Radia Perlman presented some ideas on how to handle mobile host
 (e.g., flying packet radios) based on having the gateways do most of
 the work.  This approach is based on the gateway using a "link-state"
 routing procedure and a method of handling partitioned nets as if
 each partition were a separate net.  The method presented is related
 to but not identical with that presented in IEN 120.  A revision of
 IEN 120 will be forthcoming.

 Carl Sunshine presented an alternate strategy for dealing with mobile
 hosts which makes the hosts do most of the work.  This scheme uses
 the existing IP source routing option and a new GCP message.  It
 calls for a new host (or special functions in an existing host or
 gateway) to support (1) forwarding and (2) a global name server.  New
 messages are needed in the gateway protocol (IEN 109) to relay
 information about a mobile host's current location to the name server
 and the "connected" hosts.  This procedure is described in IEN 135.
 This procedure may also be workable for multihomed hosts and as a
 least effort solution to the partition problem.

 Clearly a decision must be made as to whether these problems are to
 be solved within the gateways or not.

 Vint Cerf described another routing problem.  This problem arises
 when a destination can be reached either within a network or via
 another network which is connected to the first network in two (or
 more) places.  For example, the ARPANET and the WBCNET will be
 connected in four places.  It may be better for messages from Boston
 to Los Angeles to go one hop in the ARPANET then via WBCNET then one
 hop in the ARPANET, than cross country via many hops in the ARPANET.
 Another example, is a large Packet Radio environment surrounding a
 small ARPANET style network.  Messages from one packet radio to
 another packet radio on the opposite side of the environment might
 best be forwarded through the ARPANET style net rather than via many
 PRNET repeaters.  This out-of-net crossnet routing is a difficult
 decision to make by current procedures.  (The latter example was
 described as a "cloud of packet radios."  First we had a flying PR,
 now we have a whole cloud!  I hope it doesn't rain!!!)

XII. SOURCE ROUTING IN A CAMPUS ENVIRONMENT

 Jerry Saltzer presented his ideas on routing in a collection of local
 network in a campus environment.  Much of the strategy is based on
 the lack of central control over the environment.

 A Name Server will be provided that will supply route information to
 be used in the source route.  Source routing permits gateways to be

Postel [Page 9]

                                                               IEN 145

Internet Meeting Notes

 very simple, and allows the end user some control over the route.
 The latter is good for trouble shooting, and performance control.

 These ideas are described in IENs 143 and 144 which were distributed
 at the meeting.

XIII. CONGESTION CONTROL

 Dave Clark discussed some studies about what to do when a Source
 Quench message arrives.  A simulation program was used to try various
 methods.  All were bad.  The delay in the feedback loop is quite
 large and some information needs to be supplied about how far away
 (in time) the bottleneck is.

XIV. LOGICAL HOST SUPPORT

 Jon Postel and Bill Plummer discussed some of the issues in logical
 host support.  The problem came up with a desire to run the testing
 gateway at the same time as the regular IP.  This is not possible but
 points out the need for the IP layer to support the multiple logical
 host concept.

XV. BAKEOFF

 The "distributed bakeoff" was too distributed in space and time to be
 as effective as one would like.  Some good things did come from it.
 (E.g. see Dave Mills messages).  It was felt that the next time it
 should be scheduled for one day and with the participants in two or
 three locations.

Postel [Page 10]

                                                               IEN 145

Internet Meeting Notes

XVI. NEXT MEETING

 The next meeting will be held at RSRE in Malvern, England, on 7,8,
 and 9 October 1980.  Attendance will be restricted so if you plan to
 attend please clear that with Vint Cerf and notify Linda.  John Laws
 will be the host and information about the local arrangements will be
 distributed at an early date.  Malvern is about two and one half
 hours by train from London.

 AGENDA ITEMS

 1.  Resolution of the Partitioned Net Problem - Cerf

 2.  Proposal for Controlled Routing - Cohen

 3.  Experience with VAN Gateways - Brescia, Kirstein

 4.  Demonstration of Interim Internet Mail - Postel

 5.  Performance Evaluation Parameters - SRI

 6.  Name Server Demonstration - SRI

 7.  New ST Document and Explanation - Cohen, Forgie, Hoversten

 8.  Tiny Pipe Nets vs. the Catenet - Mills

 9.  Congestion Control - Clark

 ACTION ITEMS:

 1.  XNET Specification - Haverty, Strazisar, Mathis, Tomlinson

 2.  Tenex running TCP4 - Plummer

Postel [Page 11]

                                                               IEN 145

Internet Meeting Notes

APPENDICES: Small Group Discussions

A. FTP and Mail

 Need to extend current style of computer mail to work in the
 internet.  The extension of "Mailbox@Host" to "Mailbox@Host@Net" may
 not be acceptable because too many programs must be changed.  Rather
 something like "Mailbox@Place" is suggested.  This gives all hosts at
 least one global name for mail purposes.  (Will there be a name
 czar?)  "Place" is going to have to map to a 32-bit internet address.
 The mailer will have to try to send the mail via TCP if both source
 and destination know that or via NCP if both know that or if one is
 TCP only host and other is a NCP only host via a forwarding host that
 knows both and provides a special forwarding function.

 Postel will produce an IEN on this topic.

B. FTP

 Brief discussion of FTP and default data ports.  The data ports
 should be U and L-1 so that port pairs are not required in the users
 portspace.  Also it seems OK to use the new reply codes.

C. Measurements

 Tools are needed for measurements, for example, echo servers, and
 time stamps.  Echo servers can be TCP based, UDP based, or use the
 GGP; but more are needed.  Timestamp needs and desirable properties
 will be described in a memo by Rob Cole and Andy Bates.  Jack
 Haverty's forthcoming IEN on a traffic generator should help too.  Ed
 Cain has a report on a TCP Tester.  Also note the CMCC facilities.

D. Fault Isolation

 This discussion was subtitled "What to do when things go wrong."  The
 conclusion was "try another gateway."

E. Transport Layer Specification

 Richard Tenney is working on a task at BBN (under contract from NBS)
 to specify a Transport Protocol.  He is developing a specification
 methodology along the way.  Richard had some questions about the TCP
 specification and the actual behavior of real TCPs.  The main
 questions had to do with out-of-order data, resets, and closing
 cleanly.

Postel [Page 12]

                                                               IEN 145

Internet Meeting Notes

 Richard noted that we all should look at some NBS reports:

    "Draft Features Analysis of Transport Protocols"

    "Draft Service Specifications of Transport Protocols"

 obtainable from John Heafner of NBS (Heafner@NBS-10).

F. Fragmentation and Reassembly

 Fragmentation and Reassembly must be implemented so that datagrams of
 up to 576 octets (including header) may traverse the catenet.  It is
 proposed to hold a special bakeoff to test this capability in early
 September.

G. Gateway-Gateway Protocol

 Discussion of how much of this must be known by the hosts.  The
 messages:   Source Quench, Destination (Host/Net) Unreachable,
 Redirect, Echo/Echo Reply.  Much discussion of error reporting should
 it be in GGP or IP?  (The current IP error option seems useless).
 Postel will write a memo on error reporting (including who processes
 which errors).

H. IP Option Overflow

 The problem is that some IP header option fields (e.g., return route,
 timestamp) may expand, causing the maximum header size to be
 exceeded.  Ways to handle this might be: (1) discard the packet, (2)
 expandable options should have an "overflowed" flag which gets set
 when they can't grow any more and the packet is forwarded without
 expanding the option, (3) expandable option should be sent with
 filler so header is "right" or max length to start with.  On header
 overflow both source and destination hosts should be told out about
 the error.  The issue was not resolved.  But the problem should be
 discussed in an "Implementers Guide."

 An additional problem concerns making full length fragments when
 variable length options are present making it likely that the first
 fragment will overflow the next maximum packet size and have to be
 further fragmented later.  One suggestion is to leave the first
 fragment not quite full in such a case.

I. Name Server

 SRI will implement a name server.  It will be an extension of the one
 specified in IEN 116.

Postel [Page 13]

                                                               IEN 145

Internet Meeting Notes

J. Acknowledgement Algorithms - Adaptive Timeouts

 There seems to be a lot of ACK traffic so maybe ACKs should be sent
 on a periodic basis rather than on an event basis.  The
 retransmission strategy also needs to be smarter.  Could it be
 negotiated?  Type of service should be a consideration in these
 algorithms.  Bill Plummer and Andy Bates will conduct an experiment.
 Idea:  set retransmission time according to the Network at the other
 end of the connection.

K. ARPANET Problems

 The "8 messages outstanding at a time" problem was discussed.  It
 turns out that the maximum messages in transit can often be less than
 8 due to heavy loads at the destination IMP.  It seems that the only
 solution to this is to provide more buffers and message blocks in the
 busy IMPs.