J. Postel IEN 175 ISI 13 March 1981 Internet Meeting Notes -- 28-29-30 January 1981 I. WELCOME The meeting was held at USC Information Sciences Institute. The host was Jon Postel. Jon opened the meeting and gave the usual explanations about the arrangements. Vint Cerf extended the welcome and called special attention to the participation of representatives from CSNET, Canada, and Germany. II. OBJECTIVES Vint Cerf described the main objective of the meeting to be to put issues on the table. Problems could be resolved by subsequent arrangements between ARPA and the contractors. In particular, the topics of performance, addressing, and documentation are of special interest. III. STATUS REPORTS A. UCLA Charley Kline gave a report on the activities at UCLA. Bob Braden from the UCLA Computer Center is on a one year visit to UCL. Charley is involved in a research project in the Computer Science Department. This report is about the work Charley is involved in. UCLA is building a system called LOCUS. It is based on Unix and a ring network. The ringnet hardware is the LNI developed by UCI and modified by MIT. A UCLA developed private protocol is used on the ring. The current work is on the distributed operating system. The goal is for the users of the system and for remote hosts to see the collection of hosts and the ring net as one system. The most recent effort is on file system aspects. Very good results have been obtained yielding essentially equal access times for local and remote files. B. UCL Peter Kirstein gave most of the presentation with some help from Ron Jones. Peter noted that with Bob Braden visiting at UCL use of the UCLA 3033 from London has increased and the service is Postel [Page 1] IEN 175 Internet Meeting Notes quite good. However, he has observed random outages of line 77 lasting 5 to 15 minutes several times per week. Bob has installed echo servers both at the TCP and IP levels at UCLA for use with UCL measurements and tests. Peter described UCL's work on: (1) a protocol converter for service between SATNET and SRCNET for terminal access, SRCNET is an X.25 network; (2) a cambridge ring at UCL where terminal access is now working via UMCZ80 interfaces; (3) a local message system is up on the UCL Unix using the MMDF system developed by Dave Crocker; (4) the NIFTP will be upgraded to match the latest document; and (5) UCL has worked over the TIU TCP and made some parameter changes and installed a dynamic retransmission timeout procedure, also MOS was revised, and a document is available on this. C. RSRE John Laws gave the RSRE report. Gateway Status: The line between the UCL gateway and the RSRE gateway, known as net 35, was upgraded from 4.8Kb/s to 9.6Kb/s in November 1980. The gateway machine is still an LSI-11. Local performance measurements on the gateway were instrumented by writing a process timer for the MOS system which showed how long and how often individual processes ran for and the total time spent idling in the scheduler. These measurements showed that the gateway did handle 50Kb/s lines and would give the expected throughput on these lines for large data packets ( >128 user data bytes). The switching capability was 50 packets per second using the RSRE line cards. The time taken to switch packets from one interface to another was 4.8msecs. This includes use of hashing tables to determine the route of the packets. Although fragmentation has been implemented and tested as reported at the last meeting no use has yet been made of this facility. Source routing code awaits a final imprimatur of the doctrine to be employed. Service Availability: A comparison between ARPANET availability from the PPSN for one month periods before the last meeting and this meeting show that although the outages total a similar total time the ones for the Nov-Dec 80 period are mostly due to developments in the UCL gateway rather than to SATNET as was the case for the Aug-Sep 80 period. Two points are worth noting: (1) Some of the longer outages were due to debuging sessions on the UCL gateway. These are difficult because they require personnel at RSRE, UCL and BBN to be available simultaneously. They often require loading software at the RSRE gateway for the tests and reloading software after the tests. (2) The large number of short outages were traced to resetting of the Host-SIMP interface by the Postel [Page 2] IEN 175 Internet Meeting Notes UCL gateway caused by shortage of buffers. Ginny alleviated this to a large extent in January by removing code from the gateway. However if UCL and RSRE are heavily using the gateway we can still get up to 20 resets a day. Local Network Status: The PPSN (net 25) now has three network access protocols available to users, these are datagram type (with multi-address), virtual call (PPSN type with multi-address), and X.25 virtual call. We have developed a flexible evaluation host both for use on the PPSN and the British Post Office PSS network which we hope to be connected to in February. This evaluation host supports 2 DTEs and includes a command console for setting up multiple calls from one terminal, traffic generators, and a measurement package. The protocols employed are the Network Independent Transport Service (NITS) [formerly "Yellow Book"], and X.25 level 3 with BPO options and the RSRE X.25 level 2 line units. TCP and IP: The CORAL implementation of IP with fragmentation and reassembly is complete. The design for the CORAL TCP including full state machine representation is complete and coding is in progress. We have obviously gained a lot of useful experience for this from using the Mathis TCP. Besides TCP measurements we now have the capability to perform measurements of IP using echoers on ISIE and EDN Unix as well as GGP echo packets. The IP measurements package has the capability to make single round trip measurements at a user settable protocol type, packet length, and packet rate of 1 to 50 packets per second. Action Items from RSRE: 1. Dynamic timeouts for retransmission. 2. Flag bit for copied options in IP fragments. 3. Specification for strict source routing. 4. Standard addresses for Echo, Discard, and Character Generator servers. 5. Techniques for preventing silly window syndrome (name due to Dave Clark). 6. No changes to addressing for a while. Postel [Page 3] IEN 175 Internet Meeting Notes D. SRI Holly Nelson reported that the TIU with fragmentation and reassembly is in progress, that a new version of the port expander program was delivered, that a measurement host being installed is version 7 Unix with IP separate from TCP, and that Ron Kunzelman will be leaving SRI in mid February. E. NDRE Yngvar Lundh reported that progress at NDRE continues to be hampered by the lack of people. In spite of this, progress is being made and speech now runs in the local net. Paal Spilling has returned to NDRE after a one year visit to SRI. Paal will be trying speech experiments with Lincoln. The agreement with the Norwegian Telecommunications Authority is being renegotiated for 2-3 years. NDRE has had discussions with the German group. Action Items from NDRE: 1. A HDLC port on the SATNET gateway is needed. 2. ARPANET availability must be improved. F. MITRE Anita Skelton discussed the development by MITRE of a Z8000 based TCP. The program is written in C and cross compiled on Unix. The system is a modified C MOS. The TCP is about 600 lines of C code. This TCP has been tested talking to itself, but not against any other TCP. The data rate measured is 350 Kb/sec. Later in the meeting Steve Holmgren gave a presentation on this. G. MIT Dave Clark described developments at MIT. Planning for a computer network architecture at MIT is being formulated in terms of a system involving 52 buildings and 10,000 hosts. As Dave noted, he reported again that the version 2 ring net (10 Mb/s) is almost working. The development is a joint project with PROTEON. Some interfaces are prototyped in multiwire. The use of IP and TFTP based services is spreading. A Unix on the CHAOS net uses IP to communicate with a Unix on the Ring net. Fragmentation and Reassembly is used widely in the MIT environment. MIT did its own reworking of MOS -- this time to make it easy to bind with programs from other source languages. Postel [Page 4] IEN 175 Internet Meeting Notes On the multics front Dave reports that IP/TCP was used over a HASP connection on a 9.6 Kb/s line between Cambridge and Phoenix with a round trip time of 8 seconds. Various investigations suggest hosts should be careful about also being gateways, for example, the gateway echo traffic could be more than the host could handle. In one experiment a routing loop was detected and datagrams would have looped forever except for Multics decrementing the time to live. Multics is being connected to TYMNET and TELENET, initially as a server for remote terminal access. A NIFTP server is being developed, and a MTP server is up on both NCP and TCP for RFC 772 style mail. The TOPS20 system, MIT-XX, still has trouble with the IP implementation crashing the system, and the performance is poor otherwise. Dave put together a minimal IP/TCP/Telnet in BCPL for an Alto. The total package is about 3000 words of code. The Swallow Distributed Computing Project is basing its communication primitives on IP/UDP. A C-30 IMP is scheduled for delivery to MIT in early February. Action Items from MIT: 1. A maximum segment size TCP option is needed. H. Lincoln Laboratory Jim Forgie described the work Lincoln has done on the Voice Terminal (VT) and the ST Gateway. Several VTs now exist and are functioning on the LEXNET. Two modes of speech data are supported 64Kb/s PCM and 2.4Kb/s LPC. A ST gateway is partially implemented in an 11/45 running EPOS and experiments are being done via the ARPANET with ISI. The gateway currently supports only the point-to-point ports of the ST protocol (no conference support yet). The measured gateway throughput is about 100 packets/second. The next few milestones on the Lincoln schedule are: 1. Move the gateway from the 11/45 to the 11/44. 2. Demo speech over the WBCNET. 3. Deliver a LEXNET and VTs to ISI. 4. Add source routing to ST and the gateway. Postel [Page 5] IEN 175 Internet Meeting Notes 5. Add conferencing features. I. Linkabit Estil Hoversten discussed Linkabit's role in the support for SATNET and WBCNET. Linkabit provides clever modems to make the most of the channel. Estil pointed out that when Germany and Italy join the SATNET the channel will be more shared and less capacity will be available to existing communication partners. For the WBCNET Estil reviewed the state of installation. All the equipment is in place at ISI and LL, but the RF and power equipment needs tuning up. The DCEC and SRI equipment should be fully in place before the next IN MTG. Something called "Remote Control and Alarming" is needed to make the operation legal. The FCC requires that an operator (FCC licensed) be able to control the transmitter (shut it off) if something goes wrong. Western Union is installing the remote control mechanism so their operator in New Jersey can satisfy this requirement. Estil also mentioned that Linkabit has some communication between two sites in San Diego several miles apart to support terminal access for terminals at one site to TOPS20s at the other site. They use dumb multiplexors and a 45 Mb/s microwave link. Also Linkabit is part of M/A-COM which is planning to link its four major sites via a satellite channel in the 800 Kb/s to Mb/s range. J. ISI Jon Postel reported on the activities at ISI. The Internet Project has three areas of interest: Protocol Verification, Protocol Design, and Protocol Applications. In the verification area Carl Sunshine made a presentation later in the meeting. In the design area Danny Cohen has led our efforts primarily in discussion of addressing issues (see IEN 165). In the application area we are developing two computer mail applications: the MPM and the MTP. The MPM is a multimedia mail delivery system following the specification of RFC 759. This is now working on TOPS20, and was demonstrated at the end of the meeting. The MTP is a text only mail server for multinetwork operation and follows the specification of RFC 772. This is partially implemented on TOPS20. It can now accept mail via TCP connections and leave it for distribution via mailer and NCP connections. MTP was demonstrated at the end of the meeting. Another program was developed to better understand using TCP and can be used as a user Telnet. This program is called TT. A description was distributed at the meeting and TT was demonstrated at the end of the meeting. Other work at ISI included the development and installation of Postel [Page 6] IEN 175 Internet Meeting Notes code in the TOPS20s to guard against sending the 9th outstanding message to the same destination in the ARPANET and thus be blocked by the IMP. This has been installed in the five TOPS20s at ISI. The PDP-11 which is the front end for the Penguin printer has been modified to also route internet datagrams to other hosts on the ARPANET or the local Ethernet. The WBC project has participated with Lincoln in the Speech experiments reported on by Lincoln. Also the WBC project has interfaced a radio clock to the speech host so that the operating system (EPOS) gets the time from the radio clock. K. DoD Ray McFarland mentioned that questions are being raised about the Internet Addressing and its implications for Network Architectures, and on the relationship between Internet Architectures and Distributed Processing Systems. Later in the meeting Ken Shotting discussed his work on specification of IP. L. DCEC Ed Cain gave the report on DCEC activities. The hosts in the EDN now do reassembly. There are two gateway programs. The old gateway does fragmentation, talks GGP, and sends in CMCC reports. The new gateway performs well. There was some confusion about GGP routing messages. It seems that a type 11 message was used in the BBN-built gateways to avoid confusing a new routing message format with the one documented in IEN 109 (How to Build a Gateway). The DCEC gateway used the type 11 message in order to exchange routing information with its neighbors. However, since the agreed-upon resolution is to modify IEN 109 to show the new format, the BBN-built gateways have been gradually reverting back to the new-format type 1 message. During the transition, the DCEC gateway has kept track of which gateways use the two routing types, and serves as a routing "bridge" between the two communities. The right type code for routing messages is Type 1, and no Type 11 messages should be used anymore. DCEC also implements the security and precedence features of IP. A TIU for AUTODIN II is being developed by DCEC, and an implementation of MTP is in progress. DCEC coordinated a seminar on IP and TCP at NBS in November which 405 people attended. Postel [Page 7] IEN 175 Internet Meeting Notes Action Items from DCEC: 1. How do we provide testing facilities to companies that develop TCP products? 2. How do we control transit traffic? M. COMSAT Dave Mills gave a brief review of the configuration at COMSAT. The most important new development is the idea of virtual hosts which are processes with internet address and move from one real host to another within the DCNET. Another idea being developed is the maintenance of synchronized clocks in the hosts of the DCNET. This is done with time stamps in routing update messages between the hosts. Dave reported some interesting results about clocks and time variations that came to light in this work (see IEN 173). Also at COMSAT a MTP mail server is working. The FAX program now handles multipage documents. NIFTP works between COMSAT and ISIE -- 5 million bytes of RFCs and IENs were transferred. The INTELPOST project conversion to IP/TCP continues. A copy of the DCNET software has been installed at FORD research center at Dearborn with a 1200 baud dial up connection between Dearborn and Washington. N. BBN Jack Haverty reported on the development of three new TCPs: for the TAC, the HP3000, and the VAX Unix. In the TAC, the TCP connection opening and closing has been completed, and work continues on the data handling code. In the HP3000 version the TCP is done and work is in progress on Telnet while waiting for a hardware interface. The VAX Unix (and C70) version is passing packets but work on TCP continues. The work on the VAN gateway (ARPANET/TELENET) continues; testing of X.25 level 3 is being done with an Applied Data Communications Pro/Tester. The TIP Login design is just about done; it is being designed as a special case of resource allocation. BBN is modifying CMOS for use in the C50 environment. The CMCC data has shown some anomalous behavior. For example short periods (5-10 minutes) with a high percentage of dropped datagrams and at the same time NCC data showing very busy IMP's. Hard to track this down -- better fault isolation tools are needed. Ginny Strazisar reported on the developments in the BBN-built gateways. A new version of the gateway code was distributed to all sites (except Ft. Bragg) which will do fragmentation (but not with options). The gateway at UCL was modified to only use small Postel [Page 8] IEN 175 Internet Meeting Notes (SATNET size) buffers and remove local operator support to improve performance. The fragmentation routine will send on fragments as large as the network can handle. Work is in progress to change to a routing procedure which detects partitioned networks. Dale McNeill reported on SATNET. After several months of poor performance, on November 6 there was a significant decrease in packets received with checksum errors on the SATNET channel. Performance is consistent with a change in the BER from about 10**-4 to about 10**-6. No information on the potential cause has been given by INTELSAT. In any case, SATNET channel protocol stability and the ARPANET direct connection via SATNET performance (ARPANET line 77 to the London TIP) are much improved. The US-Norway ARPANET satellite circuit (ARPANET line 41 between the SDAC IMP and the NORSAR TIP), which was changed last spring from commercially-controlled to military-controlled, continues to provide poor service. This line is composed of a military satellite hop and a number of commercial circuits. Hence, NDRE gateway to UCL gateway traffic is still disallowed in the Tanum Satellite IMP, thereby breaking internetwork traffic between the two gateway sites. A bug in the Host-SATNET protocol was found, and a fix was installed in the Satellite IMPs. The same fix needs to be inserted in the gateway to provide complete protection in the protocol. A major milestone was passed in MATNET, the secure shipboard copy of SATNET being built for the Department of the Navy; namely, a Satellite IMP successfully transmitted packets through a Navy FLTSAT satellite for the first time. The test setup was at the plant facility of E-Systems, ECI Division, where preliminary integration is being done between the network controllers and the radio equipment. Action Items from BBN: 1. Rubber EOL and Buffer Size has implementation impact in TAC. O. ARPA Vint Cerf reported that the ARPANET switched over to 96-bit headers on January 1st. Bill Carlson is leaving ARPA to join Western Digital and plans to build ADA machines. Vint mentioned a project at CCNY which is encoding video using adaptive delta modulation. ARPA plans to make a pair of these devices available to ISI and DCEC in July 1981. ARPA plans to replace all the 316 TIPS by TACs plus C30 IMPs. TIP Login will be used for access control. Germany and Italy are likely to join SATNET soon, Canada is in the discussion stage. Postel [Page 9] IEN 175 Internet Meeting Notes P. CSNET Dave Farber gave a brief overview of CSNET. The idea is to provide network services to a group of university computer science departments. It is a 5 year NSF grant. The contractors are University of Wisconsin, University of Utah, Purdue University, The RAND Corporation, and 9 others. There is a steering committee (Chair Bill Kerns, NSF) and a technical committee (Chair Dave Farber, UDEL). CSNET will use Telenet, ARPANET, and telephones to connect the universities. The services to be provided are: Computer Mail, File Transfer, Terminal Access, and Distributed Processing Applications. More information is available from Dave (FARBER@UDEL). Q. DTI Gary Grossman gave an overview of DTI's work on a front end for the WWMCCS H6000 hosts. The INFE (interim network front end) has been up for some time, and has had some testing. It implements the January 80 IP and TCP and has been tested with the EDN systems. DTI is now working on the COS/NFE which is to be multi-level secure and will be based on DTI's HUB (TM) operating system. The COS/NFE is to be verified by SDC. R. NAVELEX Frank Deckelman described the interests and activities of NAVELEX Code 330. There are four areas: DBMS, Distributed Processing, Decision Aids, and Networking. In Networking the focus is on: Global networks (e.g., MATNET), local networks (e.g., CCN - uses Ungerman-Bass hardware), and C2 work stations (multimedia and AI). One current program that involves all four areas of responsibility is the ACCAT at NOSC and the Remote Site Modules at NPS, FNOC, CINCPACFLT, NRL, and other sites. S. XEROX John Shoch gave a brief status report on networking at XEROX. There are between 30 and 40 networks in operation, connected with 20 to 30 gateways. The new 10 Mb Ethernet is up. The XEROX series 8000 products are based on the Ethernet. There is a Print Server, a File Server, and a Communication Server (gateway), as well as work stations. John also mentioned that a paper on "Measured Performance of the Ethernet Local Network" appears in the December 1980 issue of the Communications of the ACM. Postel [Page 10] IEN 175 Internet Meeting Notes IV. ACTION ITEMS A. The problem reinitializing UCL gateway was resolved. B. The design notes on the VAX, HP3000, and TAC TCPs were done [see IENs 166,167,168]. C. The memo on IP Errors and GGP style messages was distributed in DRAFT form. D. The demonstration of the SRI Name Server was postponed indefinitely. E. The MTP was demonstrated at the meeting. V . PROTOCOL SPECIFICATION AND VERIFICATION A. Overview Carl Sunshine gave a well-received presentation on formal models for protocol analysis. Carl presented a basic model of a protocol and identified the aspects that need to be specified and verified. Then he described the methods for specifying protocols and identified particular programs or techniques that implement each method. Finally, Carl reviewed the verification methods and again identified particular systems which implement each method. The specification methods are: Programs inductive assertions State Machines FSA Abstract Machines Petri Nets Sequence Expressions The verification methods are: Program verification State Exploration Structural Induction Symbolic Execution Design Rules A DRAFT version of a report on "Formal Modeling of Communication Protocols" was distributed. The final version is ISI RR-81-89. Postel [Page 11] IEN 175 Internet Meeting Notes B. Three Way Handshake Danny Schwabe presented an analysis of the three way handshake made in the SPEX language. A description of a protocol in SPEX is a "spexification." Spexifications can be translated into an AFFIRM abstract data type. Proofs of properties of this abstract data type can be done using AFFIRM. The analysis revealed a possible problem in the three way handshake. As currently described in the TCP specification (IEN 129), a connection may become established on one side and potentially may accept data from a previous incarnation of the connection. This is a very low probability case and would be quickly detected, but the analysis showed it was possible in principle. The problem is that the document is in error in allowing an ACK to be accepted before a SYN has been received. A DRAFT report "Formal Specification and Verification of a Connection Establishment Protocol was distributed. C. IP Specification Ken Shotting described his specification of IP using SRI's HDM and SPECIAL. Ken broke down IP into several mini-layers to make use of the methodology. There was some difficulty making assertions about global behavior since several fields are changed between the source and destination, e.g., Time-to-Live. This also causes problems with the checksum field. The order of processing assumed for various features has an impact on the formal specification. Ken's evaluation of HDM for protocol analysis is that it provides good support for developing a hierarchy of abstract functions and for state machine transitions, but is weak in data representation, exception handling, and concurrency. Ken's work is documented in a University of Maryland Technical Report "On the Formal Specification of Computer Communication Protocols," Computer Science TR-973, December 1980. D. NBS Work Jack Haverty gave an overview of the work BBN is doing for the NBS. The work is being done by Tom Blumer (TPB@BBN-Unix) and Richard Tenney (RTenney@BBN-Unix). This protocol specification system is based on a finite state machine model, augmented so that there are variables, and arbitrary program segments may be associated with each transition. These program segments are written in a Pascal-like language. There is a syntax checker for specifications, a specification editor, a FSM analyzer, a compiler that produces C and a test generator. Postel [Page 12] IEN 175 Internet Meeting Notes For further information contact Tom Blumer or see BBN Report 4581. E. DCA Work Dave Kaufman reported on the work SDC is doing for DCA. The main emphasis is on developing complete specifications of existing protocols based upon the original design intent and upon a set of specification guidelines being prepared in parallel by SDC. These guidelines include the following sections: Introduction, Service Specification, Lower Level Service Specification, Interface Specifications, Peer Level Mechanism Specification, and Execution Environment. These sections correspond with the basic protocol model presented by Sunshine. A distinction is made between the service specification which is the global view of the service and the interface specification which is the user/service local interface. The execution environment section describes protocol requirements of the operating system (or more correctly of the protocols runtime environment) such as timers and interprocess communication. IP is currently being specified according to these guidelines and during this effort several areas have been uncovered which require further definition. Example areas noted by SDC were: 1. Who supplies the "protocol" field for the IP header, IP or the transport protocol? 2. What is the nature of the interaction between IP and GGP? 3. Is source routing loose or strict? Jack Haverty raised an additional related issue: 4. How does IP interact with the local net, on errors, on flow control,etc.? F. NDRE Work Yngvar Lundh noted that a graphical technique had been used to produce an augmented state machine description of TCP for use in an NDRE implementation effort. Postel [Page 13] IEN 175 Internet Meeting Notes VI. FRONT ENDS A. Overview Vint Cerf introduced the subject of Front Ends for protocols and TCP in particular. The idea seems to be that in some cases IP/TCP etc. may be too complex to put into the existing operating system, so the protocol modules are put into a small front end machine. Then there must be a communication procedure between the host and the front end. This communication procedure is called Host-Front End protocol (HFP). The argument is that HFP is much simpler for the host than IP/TCP would be. Note that not everyone shares this view. B. WWMCCS HFP Gary Grossman gave a detailed presentation on the HFP developed for the WWMCCS H6000 machines. DTI has developed a protocol for use between the host and the front end based on the notions of "services" and "channels." The lowest level of the HFP is the link level which provides a single channel with flow and error control. The function of the link level corresponds to the functions of an HDLC connection. The middle level is the channel level. This is the level where separate logical streams appear and are multiplexed based on logical channel number. At the top level are services (i.e. applications). Typical services would be Telnet and FTP. Gary gave quite a bit of detail on the protocol formats and control procedures. Gary also talked about the specification technique used to describe the HFP. For each message type the following points are described: function, when sent, sending state (and next state), receiving state (and action to take), subsequent actions by sender and receiver. For the service protocols, a specification includes a "specification" and an "adaptation." The specification gives the semantics of the fields used and the procedure for communicating via the channels. The adaptation gives the format details and any restrictions due to the particular machine environment. DTI and MITRE have done a fair amount of performance testing. The comparison of an old implementation of NCP in the H6000 versus an NCP implementation in the Front End show the Front End about 2 to 1 better in throughput. There are many differences in the two configurations. Some testing recently showed that from the H6000 through the FE and over the ARPANET to DTI a data rate of 18Kb/s was achieved. In local tests, 64Kb/s was measured over the H6000-->FE-->I path; and with a source and sink in the FE, 84Kb/s Postel [Page 14] IEN 175 Internet Meeting Notes was measured for source-->I-->sink. Finally using a "mirror" program in the FE instead of sending to the IMP, 125Kb/s was measured for source-->mirror-->sink. An analysis of the CPU time in the TCP code in the FE revealed that about 20% of the time was spent on TCP functions and 80% on interprocess communication and other "system" functions in the monitor. Gary noted that details of the HFP are described in DTI report DTI-80043.C-INFE.18, and DTI-78012.C-INFE.14, and the measurements are described in the papers "Control Structure Overhead in TCP," NBS Computer Networking Symposium, May 1980, and "A Performance Study of a Network Front End," Sixth Data Communication Symposium, November 1979. C. Z8000 Steve Holmgren of MITRE described a Z8000 based TCP. The goal is to provide network communication for very small hosts which might otherwise be peripheral devices of a computer. The Z8000 actually supports a Network Access Protocol (NAP). The NAP is a layered protocol with an option approach. The layers are: link, transport, service, and user. The option approach allows the users of the protocol to select the features they need and to ignore others. Steve gave some details of the protocol functions and formats. The access to the protocol is via a system call style mechanism, and the user is allowed to pass parameters and select options. D. Discussion John Shoch said he questioned the desirability of Front Ends, and wondered, if aside from communication efficiency and pragmatics, there was a good technical principle for front ends? It was noted that we often excuse the current work of computing checksums by saying we will someday have a checksum instruction in our machines. Steve Holmgren asked "Why not a TCP instruction?" When does a "front end" become an "instruction?" Postel [Page 15] IEN 175 Internet Meeting Notes VII. PERFORMANCE A. Multics Dave Clark discussed his measurements of TCP performance in Multics. Dave described some detailed timing studies made of the different TCP functions. The IP implementation is 5.3 K instructions, and the TCP implementation is 9.0 K instructions. Dave also discussed throughput in TCP when there is a lot of data to send (e.g., in file transfers). A problem, previously pointed out by Dave Mills, can arise where many small segments are sent. This is due to the receiver reporting additional window each time a small amount of the data has been processed, and the sender immediately sending a segment to fit that additional window. Dave Clark proposes to call this phenomena "Silly Window Syndrome" (SWS). One way to prevent SWS is for the receiver not to report new window unless the increase is a reasonable size. The receiver can acknowledge incoming segments at any time, but limit window updates to points when a reasonable increase can be made. It is up to the receiver to decide what is reasonable, perhaps something like 20% of the total buffer space for this connection. The sender can also try to stay out of SWS by only sending big segments and waiting until the window is large enough to allow it. If the sending user indicates an end of letter then a smaller segment may be sent. Dave suggests that sending a segment with one octet of new data into a zero window as a probe may lead to SWS. It may be better to send an octet of old data. Dave suggested that a performance measure might be the size of bursts of segments the net, gateway, or host could handle. For example, can a gateway handle a burst of 10 datagrams of 576 octets at the network input speed? Could measures be devised and feedback control be exercised in terms of bursts? Wild idea number one: the receiver can tell the sender the maximum burst size it should send. Wild idea number two: have gateways turn on a bit in the IP header if they are busy, and have the receiving hosts integrate this information for use in determining a burst size to feedback to the source. B. UCL Ron Jones reported on measurements of datagram traffic between UCL and ISIE. The source of the traffic was an LSI-11 host. This was connected to a port expander. The PE was in turn connected to the UCL Gateway. The UCL Gateway communicates via the Goonhilly SIMP, Postel [Page 16] IEN 175 Internet Meeting Notes the SATNET, and the ETAM SIMP, to the BBN Gateway. The BBN Gateway sends the datagrams through the ARPANET to ISIE. Ron had a number of measurements which he described, but interest focused on the discovery that in some tests from 25% to 90% of the datagrams were lost between the ETAM SIMP and the BBN Gateway. Some of the tests showed a significant step in the throughput graph at the single-packet/multi-packet threshold. C. BBN BBN believes that the problem originated because the ARPANET IMP 40 would not accept packets from the BBN gateway fast enough. The packet loss is manifest as multiple packet retransmissions from the Etam SIMP to the BBN gateway, until the packet eventually times out and is discarded in the SIMP. This behavior underscores two inadequacies in the system. First, the BBN gateway should discard the packet and send the appropriate message to the Etam SIMP; the machinery is already present in Host-SATNET protocol to accommodate this. Second, the Etam SIMP should notify the Goonhilly SIMP which should notify the UCL gateway that problems are arising. Unfortunately, a flow control mechanism like source quenching was never developed within SATNET, although its need has been known for some time; hence, when packets are dropped in SATNET, gateways are never informed and therefore cannot generate internet source quenching messages. D. RSRE John Laws presented some findings of measurements performed by Andrew Bates. These are datagram measurements of a single round trip (at earlier meetings RSRE has reported on TCP measurements of double round trip times). The source/sink is a TIU connected to the RSRE gateway. One set of measurements seems to show that the round trip time to the ETAM SIMP is about 2 seconds with very little spread, but the round trip time to the BBN Gateway is also about 2 seconds but with about 25% of the datagrams delayed (a bimodal Histogram). The measurements were made at different times of day, and may be due to a difference in the load on SATNET. E. CMCC David Flood Page told of tracking down the cause of some looping datagrams. In December the CMCC data showed that in the course of a day one of the PRNET/ARPANET gateways had processed several million datagrams. When the people who normally use that gateway were asked what experiment they were conducting, they replied "What experiment? We aren't doing anything." Several instances Postel [Page 17] IEN 175 Internet Meeting Notes of this incredibly high traffic level in at least two different PRNET/ARPANET gateways were detected over the next several weeks. David began investigating and eventually found the cause. The loader server that loads the port expanders attached to the PRNET gateways uses XNET 2.5, with Internet 2.5 headers, to do the loading. The loader server sends the packets to logical host 15 (decimal) on the port expanders ARPANET address, which is the address recognized by the bootstrap (obviously). The last packet sent by the loader server is the one that says "go" to the bootstrap; this causes the bootstrap to send out an acknowledgment and then start up the port expander software. One of the first things the PE does is to flip the ready line to the IMP. Now if the acknowledgment from the bootstrap is still in the IMP, and the IMP sees the ready line down, it will drop the ack (and any other outstanding packets from that host). So the loader server gets no ack, and retransmits the "go" command. The "go" command arrives at the port expander addressed as before, i.e. to logical host 15 on the PE; but the PE doesn't have a logical host 15. It is set to hand off any packets which it doesn't know how to route to its logical host 0, which is the gateway; the gateway sees that it is a packet destined for somewhere on the ARPANET that is not itself, so sends it back out its ARPANET interface, i.e. to the PE. The PE has the same trouble as before, so we get a loop. Because the packet is an Internet 2.5 packet, it does not have a time to live field, so the packet loops indefinitely until either the PE goes down, the gateway goes down, or a flood of real packets causes enough congestion to result in the dropping of the looping packet. This didn't happen every time the PE was loaded, because some of the time, the acknowledgment got out of the IMP fast enough not to be dropped when the ready line was flipped. It was largely a matter of luck. The problem will disappear either when the new gateway version, which drops all Internet 2.5 packets, goes into the PRNet gateways; or when the V4 bootstrap goes into the PEs. In fact I think the new gateway version may already be in place. Holly did put a fix in the PE as well, but I'm not sure what it was. One important point to note is that the port expander is on the ARPANET side of the gateway, not the PRNet side. David says the lesson that should be learned from this is that more remote access to debugging tools is needed. The key piece of evidence in solving this problem was provided by a "packet Postel [Page 18] IEN 175 Internet Meeting Notes printer" trace program that could only be run locally at the gateway/port expander site - which is normally unmanned. F. Radio Clocks Steve Casner described (actually showed) one of the Radio Clocks. Since we are about to start doing coordinated measurements of one way times we must have an agreement on what size and precision of time stamps to carry around in datagrams and programs. The "no thought" proposal is 32 bits of milliseconds since 1 January 1980. Others suggested that we might want more precision later so maybe we should have micro seconds and more bits. This was not resolved. VIII. WORM PROGRAMS John Shoch briefly described an experiment in distributed computing. Programs were constructed to operate in segments with each segment allocated to a personal computer. The total program is called a worm. The basic version of the program simply maintains its existence. If a segment dies the remaining segments cooperate to find a free machine, load it with the segment code, and start it running as part of the worm. John described some of the interesting things which can go wrong, and ways to prevent them. The experiment is discussed in more detail in IEN 159. One comment to the internet group is that this experiment showed the usefulness of multidestination or group addresses. The conclusion is that with datagram style communication, high speed local networks, and new control procedures, the tools are in hand to begin some very interesting multi-machine applications. IX. IP ERRORS Jon Postel distributed a draft memo on error reports in IP. The memo is titled "What every host should know about GGP". The intention is to use the GGP messages on routing advice and destination dead as a base and to add a few other messages to cover the host-to-host error conditions. The discussion focused on whether or not this should remain part of GGP or be in a separate protocol. From an implementation point of view there seems to be little difference, from an administrative point of view it seems to be cleaner for it to be separate. There were some strong difference of opinions. The matter was not resolved. Postel [Page 19] IEN 175 Internet Meeting Notes X. ADDRESSING DISCUSSION A. Overview Vint Cerf opened the topic with a call for a very general discussion of the issues and spent some time developing a list of goals for the the addressing and routing procedures. The goals were: 1. Keep routing simple. 2. Keep routing tables small. 3. Keep computation costs small. 4. Most datagrams should be delivered. 5. Use any available connectivity. 6. Multidestination delivery. 7. Handle mobile hosts. 8. Efficiently use constituent networks. 9. Support multi-homing. 10. Dumb hosts should be included. 11. Pantheism should be dealt with. 12. Degradation should be automatically fixable. 13. The system should scale up. 14. The system should be measurable. 15. Ability to use hierarchy. There was some discussion of these points and it was noted that some are service specifications while others are "good job" criteria. B. Presentation Noel Chiappa described the MIT complex which includes 9 networks and three major protocol families (IP, PUP, and CHAOS). All these networks share one "Internetworkly known" network number. Noel Postel [Page 20] IEN 175 Internet Meeting Notes listed a number of problems and made some observations. One possible long term development may be that all hosts are on local networks and only gateways are attached to long haul networks. The thrust of Noel's remarks seems to be that things are going to grow in a complex, but generally hierarchical, way; and any workable system will have to be prepared to grow, probably by taking advantage of the structure. C. Discussion Vint Cerf led a further discussion on addressing. The main focus was on the tradeoff between a flat address space and a hierarchical one. In a flat address system there is no relation between the address and the location, and routing has to be by central knowledge or broadcast. In a hierarchical address system the address is composed of fields which identify the location in a routing tree. Vint suggests that we have both in one! Let an address be composed of two parts: a hierarchical address (called an address) and a flat address (called an identifier). The address can be used as a routing guide or hint, but the identifier must match for a message to be delivered. The identifiers are globally unique names for hosts and do not change when hosts move. There are many questions to be answered in this scheme. For example, How does source routing work? Or Multi-homing? XI. FILE TRANSFER IMPLEMENTATION STATUS Vint Cerf took a survey of the implementers present to find out the status of FTP implementations. There are four different FTP systems being used: ARPA FTP (RFC 765), AUTODIN II FTP (IEN 101), NIFTP (IEN 99), and TFTP (IEN 133). The first three use TCP, and the last (TFTP) uses UDP. ARPA FTP DCN (COMSAT) now TOPS20 (BBN) April 81 EDN-Unix (DCEC) June 81 Multics (MIT) after June 81 AUTODIN II FTP BBN-Unix (BBN) now EDN-Unix (BBN) now Postel [Page 21] IEN 175 Internet Meeting Notes NIFTP TOPS20 (UCL) now DCN (COMSAT) now Multics (MIT) June 81 UCL-Unix (UCL) now TFTP Multics (MIT) now TOPS20 (MIT) now ALTO (MIT) now Unix (MIT) now TOPS20 (ISI) March 81 EPOS (ISI) April 81 XII. COMPUTER MAIL A. Mail Facilities Peter Kirstein discussed the problems of providing mail services between users in the UK and the US. Within the UK, computer mail will almost certainly be based on NIFTP and will likely be as proposed in IEN 169. One major concern for the US/UK transmission is cost. It is quite important that a minimum number of transmissions be made and especially that only one copy of a message destined for multiple users be sent across the Atlantic. Peter raised the issue of an interface between NIFTP mail and MTP mail, since it seems that the Internet will use the MTP mail. B. Mail Transfer Protocol Jon Postel discussed the MTP mail plan and indicated that a revised document would be issued soon which will describe the MTP in a network independent style. Jon agreed with Peter that a NIFTP-mail/MTP interface data structure should be defined soon, and promised to do so. There are several working (or partly working) MTPs already: Multics, COMSAT, DCEC, and ISI. C. TELEX/TWX Gateway Geoff Goodfellow described a system that is under development at SRI to connect Telex and Computer mail. A user at a computer terminal runs a program like SNDMSG or MM and simply adds some Postel [Page 22] IEN 175 Internet Meeting Notes additional fields to the header of the message to give the information needed for sending a telex. The message is routed to a special host which checks the telex information and if it is acceptable sends a telex. When a telex arrives at SRI it is processed by the special host and if the required keywords and information are found it is sent as a RFC 733 style message to the recipients computer mail mail box. A small percentage of the incoming telexes must be handled by a human operator because they do not contain the required information in a computer parsable form. D. Discussion There was some discussion of mailbox addresses and the problem of sending (and answering) mail to (from) mailboxes in other systems (e.g., Internet, TELEMAIL, ONTYME). XIII. NEXT MEETING The next meeting will be held 1-2-3 June 1981 at COMSAT in Washington D.C. Dave Mills (Mills@ISIE) is the host. Please send agenda items to Jon Postel (Postel@ISIF). If you plan to attend please register with Linda (Linda@ISIF). Local arrangements information will be sent only to those registering. The fall meeting will be held at UCL on 21-22-23 September 1981. The winter meeting will be held at SRI in mid January 1982. Postel [Page 23] IEN 175 Internet Meeting Notes DOCUMENTS DISTRIBUTED Author Subject Number ------ ------- ------ Cerf INFOCOM82 Call for Papers --- Postel AGENDA --- Postel Recent Document List --- Postel IEN INDEX --- Postel Assigned Numbers 776 Farber CSNET Description --- Postel TT --- Chase TTLUSR --- Postel What Every Host Should Know About GGP --- Cohen On IP-Addressing 170 Cohen Addressing in the ARPANET, another visit 171 Sunshine Formal Modeling of Communication --- Protocols (DRAFT) Schwabe Formal Specification and Verification of --- a Connection Establishment Protocol (DRAFT) Bennett A Simple NIFTP-Based Mail System 169 Postel [Page 24] IEN 175 Internet Meeting Notes RECENT DOCUMENTS Author Subject Number ------ ------- ------ IENs Shoch Notes on the "Worm" Programs - Some Early 159 Experience with a Distributed Computation Postel Internet Meeting Notes - 7-8-9 October 1980 160 Jones A Proposal for Simple Measurement Support 161 for Users Through Slow Nets Pershing Transport, Addressing, and Routing in the 162 Wideband Net Jones Echo Delay Measurements with GGP Packets 163 Stern CMOS System Overview 164 Cohen About Addressing in the WBnet 165 Hinden Design of TCP/IP for the TAC 166 Sax HP3000 TCP Design Document 167 Gurwitz VAX-UNIX Networking Support Project 168 Implementation Description Bennett A Simple NIFTP-Based Mail System 169 Cohen On IP-Addressing 170 Cohen Addressing in the ARPAnet, Another Visit 171 Mills Time Synchronization in DCNET Hosts 173 RFCs Postel Internet Protocol Handbook 774 Table of Contents Mankins Directory Oriented FTP Commands 775 Postel Assigned Numbers 776 Postel [Page 25] IEN 175 Internet Meeting Notes ATTENDEES Name Affiliation Nationality Mailbox ---- ----------- ----------- ------- Vint Cerf ARPA USA Cerf@ISIA Robert Kahn ARPA USA Kahn@ISIA David Flood Page BBN British DFloodPage@BBNE Jack Haverty BBN USA JHaverty@BBND Robert Hinden BBN USA Hinden@BBNE Charles Lynn BBN USA CLYNN@BBNA Dale McNeill BBN USA DMcNeill@BBNE Paul Santos BBN USA Santos@BBNE Virginia Strazisar BBN USA Strazisar@BBNA Gerry Amey Canada-DND Canadian Amey@ISIA Hoi Chong COMSAT Rep. China Chong@ISIE David Mills COMSAT USA Mills@ISIE Marvin Solomon CSNET USA UWVAX!Solomon@BERKELEY Ed Cain DCEC USA Cain@EDN-UNIX Wayne Shiveley DCEC USA Shiveley@BBNB Hans Dodel DFVLR Germany --- Helmuth Lang DFVLR Germany --- Gary Grossman DTI USA grg@DTI Horst Clausen IABG Austria Clausen.IABG@Mit-Multics Ray McFarland DoD USA McFarland@ISIA Ken Shotting DoD USA Shotting@SRI-kl W. Bradbury I.P. Sharp Canadian --- Stephen Casner ISI USA Casner@ISIB Danny Cohen ISI USA Cohen@ISIB Randy Cole ISI USA Cole@ISIB Dan Lynch ISI USA Lynch@ISIB Bill Overman ISI USA Overman@ISIF Jon Postel ISI USA Postel@ISIF Danny Schwabe ISI Brazil Schwabe@ISIF Carl Sunshine ISI USA Sunshine@ISIF Estil Hoversten Linkabit USA Hoversten@ISIA Jim Forgie LL USA Forgie@BBNC Noel Chiappa MIT British JNC@XX Dave Clark MIT USA Clark@MIT-Multics Steve Holmgren MITRE USA Steve@MITRE Anita Skelton MITRE USA Anita@MITRE Frank Deckelman NAVELEX USA Deckelman@ISIA Oyvind Hvinden NDRE Norway Oyvind@DARCOM-KA Yngvar Lundh NDRE Norway Yngvar@DARCOM-KA Merle Neer NOSC USA Neer@ISIA Mike Wahrman RAND USA Mike@RAND-UNIX Derek Barnes RSRE British T45(ATTN.Barnes)@ISIE John Laws RSRE British T45@ISIE Dave Kaufman SDC USA Kaufman@ISIE Postel [Page 26] IEN 175 Internet Meeting Notes Geoff Goodfellow SRI USA Geoff@SRI-ka Ron Kunzelman SRI USA Kunzelman@SRI-KL Jim Mathis SRI USA Mathis@SRI-KL Holly Nelson SRI USA HNelson@SRI-KL Zaw Sing Su SRI Canada ZSu@SRI-KL Ken Biba SYTEK USA Biba@SRI-KL Ron Jones UCL British UKSAT@ISIE Peter Kirstein UCL British PKirstein@ISIA Charles Kline UCLA USA CSK@UCLA-Security Dave Farber UD/CSNET USA Farber@UDEL John Shoch Xerox USA Shoch@PARC Postel [Page 27]