From braden Wed Nov 30 13:30:55 1988 Received-Date: Wed, 30 Nov 88 13:30:55 PST Received: from braden.isi.edu by venera.isi.edu (5.54/5.51) id AA00484; Wed, 30 Nov 88 13:30:55 PST Date: Wed, 30 Nov 88 13:30:43 PST From: braden Posted-Date: Wed, 30 Nov 88 13:30:43 PST Message-Id: <8811302130.AA02852@braden.isi.edu> Received: by braden.isi.edu (5.54/5.51) id AA02852; Wed, 30 Nov 88 13:30:43 PST To: end2end-interest Subject: Minutes of E2E TF Meeting Status: R END-TO-END TASK FORCE Minutes of the 11th Meeting August 16 at Stanford University MINUTES TAKEN BY: Bob Braden 1. INTRODUCTION Since many of the End-to-End Task Force members were coming to Stan- ford for SIGCOMM '88, we held a one-day meeting at Stanford the day before SIGCOMM. This meeting substituted for a video teleconference originally scheduled for June 29. Present: Unable to Attend: Lorenzo Aguilar (SRI) Gerd Beling (FGAN) Bob Braden (ISI, Chair) Eric Cooper (CMU) Dave Cheriton (Stanford) Joel Emer (DEC/MIT) Dave Clark (MIT) Eric Cooper (CMU) Jon Crowcroft (UCL) Steve Deering (Stanford) Van Jacobson (Berkeley) Bill Nowicki (SMI) 2. STATUS REPORTS Dave Clark -- MIT John Romkey took Van's new TCP, removed all Unix-specific system interfaces (including mbuf's), and counted instructions in the main line. The total was 190 instructions, in addition to per- byte operations. This implies the theoretical possibility of TCP running at 800Mb/s on a SPARC chip. Tim Shepard has a program to produce "Van Jacobson" diagrams. He now wants to automate the analysis of such diagrams. Clark's network simulator has been sent out to 6-10 sites. A color display is required to run it. The person who maintains the code is ATHEYBEY@ptt.lcs.mit.edu. It has been ported to a Sun under X.10/X.11 by Paul Schrager at UDEL. [Page 1] END2END Meeting Minutes Stanford U., August 1988 Dave Cheriton -- Stanford VMTP is now running on a Sun under Sun OS 4.0. Eric Nordmark is coming back for 6 months and will continue working on it. Cheriton is trying to get into production with the NAB board, which "knows" the VMTP packet format. He suggests that a general-purpose I/O board could be built if we could standardize on the format and location of the checksum and authentication information in all transport protocols. There was a spirited dis- cussion of the trade-offs in checksumming onboard or outboard. Dave Clark asked: how fast can a SPARC chip compute the checksum? Cheriton has defined a new VMTP service interface, a description of which he distributed. Steve Deering -- Stanford BBN's multicasting routing daemon based on RIP is up and running. Their code can act either as an IP gateway with multicast routing, or as a multicasting agent that tunnels via encapsulation. Van Jacobson talked to Mike Karels and settled the multihoming issue for the system interface to IP multicasting. The inclusion of an ISO stack in BSD is forcing much redesign of the BSD net- working code; during this process, Van is putting in hooks for IP multicasting. John Moy wants to use IP multicasting (on the local network) in the open SPF IGP protocol that his IETF working group is develop- ing. Bill Nowicki -- Sun Microsystems NFS: Off-again, on again. Sun has again changed their mind on NFS, resurrecting the earlier Version 3 that was purported to solve all the problems that our TF raised with Version 2. Rusty Sandberg [who met with the END2END TF in Jan 87 to talk about NFS] is working on NFS implementation again. Bill will distribute copies of the Version 3 spec when it is again ready. Bill has been adding timers to OS 4.0 NFS, using Van's RTT estima- tion algorithms. With the changes, he has been able to success- fully mount the RFC directory at SRI-NIC. Bill has also been experimenting with dynamically choosing page sizes, but finds that it dies when there is a slow server. Van [Page 2] END2END Meeting Minutes Stanford U., August 1988 commented that the stability diagram [what's that? RTB] predicts disaster for this. Bill thinks the Internet needs an architected mechanism to establish the bandwidth (rate) as well as the MTU over a given path. Van suggested that the way to do round-trip timing is to put a timestamp into a packet [as RFC-1072 "TCP Extensions for Long- Delay Paths" suggests for TCP]. There was a discussion of rate control, and the problem that the interval that is needed to avoid overruns is smaller than the clock resolution (20ms). Cheriton pointed out that the NAB will supply a fine interpacket interval, and that what is needed is a similar facility in a general Ether- net chip. It was also noted that MicroVax II's have a microsecond CPU clock. He mentioned that AT&T wants to run NFS over TCP instead of UDP. Crowcroft -- University College London Jon will [did] present a SIGCOMM '88 paper on a reliable multi- casting extension to the UCL Sequenced Exchange Protocol (ESP). He pointed out that the same scheme could easily be adapted to TCP, by binding the multiple endpoints of the communication at SYN (connection establishment) time, and building a PCB for each end- point as seen from a particular host. The windows are all cou- pled. A 1<->n multi-connection would reduce the network traffic by a factor of 2 relative to having n parallel TCP connections (assuming IP multicasting is available for sending data). [At this point, Van's eyes lit up, and we got the feeling that multi-connections might magically appear in some future release of of Berkeley TCP.] UCL is actively experimenting with a working FDDI network, using MAC-level bridges. UCL has been performing RPC experiments from LISP over ESP. UCL is involved with development of the RACE Broadband ISDN [my notes say "ATM", but I don't remember what this referred to]. This is a cooperative research effort among European PTT's. The plan is to start to deploy a real network across Europe using this technology in 1992. Dave Clark observed that the SMDS effort in the US, which is pointing towards DS3 service, is to be working in 3 years. The model will be an SMDS network with 20,000 packet- switching gateways around the outside. Finally, UCL is continuing to work on ISO protocols. They are [Page 3] END2END Meeting Minutes Stanford U., August 1988 developing an X.500-to-DNS(bind) relay. He predicts that X.500 will put the current Internet DNS out of business. Craig Partridge -- BBN As Steve mentioned, Waitzman has the multicasting router for BSD running, using a Loose Source Route trick to tunnel multicast datagrams across the Internet. Waitzman is working on an RFC describing the modified RIP for multicast routing. It uses "trun- cated reverse path broadcasting". Craig noted that BBN Labs has become BBN Systems and Technologies. Craig has implemented the MTU Discovery IP option in Van's new BSD routing code. He has found that the cost to a host for this algo- rithm is high. Van Jacobson -- Berkeley VAN has been working on pushing TCP speeds, and has reached the point where the fundamental limitation is the memory bandwidth (~20Mbps). Dave Clark suggested that Van try using an Ardent Titan instead of a Sun; the Titan's bandwidth is ~ 1.2Gbps! Van noted that IBM is supplying him with RT's, so he can add his performance improvements to increase NSS performance by *2. Van discussed TCP performance on an Ethernet, which is dominated by interference between data and ACK packets, mostly colliding at the data receiver. About 10% of the bandwidth is lost due to the ACK's. He showed a typical graph: |Throughput | | * | * * | * * * | * | * | * #packets/buffer -> |*_____________________________________________ Clark asked why the receiver always loses. It is because the sender holds the wire for a relatively long time while sending a big packet; as soon as it finishes, everyone else tries to get [Page 4] END2END Meeting Minutes Stanford U., August 1988 service and collides. Van has a bunch of analysis tools, all X-windows based. This remark led to the suggestion that the Task Force hold a Laboratory Day at the next meeting, which is to be at MIT. Van will bring his goodies and show them off. An undergraduate under Van's direction wrote a public-domain TCPDUMP, using a new kernel-resident packet filter that has some IP-specific pseudo-instructions. He also has a compiler that pro- duces instructions for the filter abstract machine. Van described a tool, originally due to the mathematical statisti- cian Tukey, that allows one to display data with the time scale reduced module T. On a workstation one can adjust T until the display merges, which provides a very powerful tool for looking at network data. [My notes mumble something about the ideal gas law here ???] Van pointed out that for a LAN or for any streaming application, you don't want to use the Nagle algorithm, since it stops pipelin- ing. Need to be able to turn it on or off depending on the RTT. Satya at CMU has a reliable multicast for file distribution. Craig will ask Waitzman to investigate. Finally, Van is looking at building a transaction protocol on top of TCP by sending a segment with SYN, FIN, and Data (which was called a Kamikazee or Christmas Tree packet during TCP/IP research days). It would be sent using the BSD sendto() call. Clark pointed out that to make this work a TCP must keep state (the last sequence numbers). Bob Braden -- ISI Bob has finished up and distributed NNStat, a package for distri- buted collection of long-term network statistics. The data acquisition element is a program named statspy that can be dynami- cally configured to collect a wide variety of statistics from Eth- ernet packets gathered promiscuously. It runs on Suns under SunOS 3.x. Bob has started a major task of writing and editing the Host Requirements RFC, that will describe the Internet current archi- tecture for hosts in great detail. He serves as chair of the Internet Engineering Task Force Working Group that is concerned with creating this document. [Page 5] END2END Meeting Minutes Stanford U., August 1988 3. STATE OF FUNDING Dave Clark made some comments on the state of network research fund- ing. 4. MULTICASTING Craig Partridge pledged to get the code being developed by BBN into the research community [he delivered on this pledge: RTB]. He noted that Waitzman's code needs more work. Mike Karels wants to use multicasting for many local applications. There are also commercial customers... e.g., Reuters wants multicast. [Let's see, VMTP is controlling a pulp mill in Finland, while IP mul- ticasting will be used to distribute news stories... We are having quite an impact on the world.] Clark asked: what multicast routing algorithm should be used in the MIT gateway? Steve answered: that depends; you have a choice of RIP- or SPF-based routing algorithms. Steve gave advice on how to install multicasting on the MIT network and gateway. Clark noted that sources for the MIT gateway are freely available and that he is trying to interest a funding agency in using MIT gateways as a platform for gateway experiments. MIT already plans to add SNMP and the SPF OIGP to the gateway. On a MicroVax III it runs at 1200 packets/sec. 5. NEW ACTION ITEMS No new action items were developed. 6. NEXT MEETING The next meeting will be held at Dave Clark's Laboratory for Computer Science, MIT on November 3-4, 1988. [Page 6] END2END Meeting Minutes Stanford U., August 1988 [Page 7]