From iso-RELAY@SRI-NIC.ARPA Sat Mar 25 11:12:47 1989 Posted-Date: Sat, 25 Mar 89 10:57:43 PST Received-Date: Sat, 25 Mar 89 11:12:45 -0800 Received: from SRI-NIC.ARPA by venera.isi.edu (5.61/5.51) id AA18991; Sat, 25 Mar 89 11:12:45 -0800 Received: from nrtc.nrtc.northrop.com by SRI-NIC.ARPA with TCP; Sat, 25 Mar 89 10:59:09 PST Received: from ames.arc.nasa.gov by nrtc.nrtc.northrop.com id aa06196; 25 Mar 89 10:59 PST Received: by ames.arc.nasa.gov (5.61/1.2); Sat, 25 Mar 89 10:58:19 -0800 Received: from hamlet.ultra.com by ultra.com (3.2/SMI-3.2) id AA15777; Sat, 25 Mar 89 10:56:59 PST Received: by hamlet.ultra.com (3.2/SMI-3.2) id AA01651; Sat, 25 Mar 89 10:57:43 PST Date: Sat, 25 Mar 89 10:57:43 PST From: Bob Beach Message-Id: <8903251857.AA01651@hamlet.ultra.com> To: iso@nrtc.northrop.com Subject: Re: TCP/IP vs. OSI Performance Status: R This may be surprising to most people but the fastest transport implementation around is based on ISO TP4. We at Ultra Network Technologies run TP4 in excess of 102MBytes/sec (i.e. 816Mbits/sec) with checksums ON. This is using class 4 transport and the inactive NL subset over a DL that operates at 125Mbytes/sec (1Gbit/sec). The connection is between a Cray 2 and an Ultra developed frame buffer. We can operate between Crays in excess of 66Mbytes/sec. In both situations we are limited by the rate at which we can get data out of and into the Cray, not by our implementation. We use their HSX channel (105 Mbytes/sec) and an Ultra developed transport engine that is attached to the channel. We impose virtually no load on the Cray (except for moving data in and out of the main memory). We also have connections to minisupercomputers (Convex and Alliant) and workstations (Suns, SGI). We operate between minisupers in excess of 13Mbytes/sec (out of 15Mbytes/sec possible) and between Suns in excess of 4.5Mbytes/sec (out of 6Mbytes/sec possible). We looked at TCP/IP and decided we could run ISO faster because: 1. TP4 headers are much easier to parse than TCP/IP's. Compared to TCP's, the DT and AK headers are almost trivial. 2. TCP's sequence number/credit mechanism is based on bytes, whereas TP4's is based on TPDUs. This allows us to open the window much wider under TP4 than under TCP. For high speed nets, its also seems "nicer" for the basic unit of information to be a packet rather than a byte. I agree with the general comments about OSI, especially at higher layers but at TL and below, ISO seems provide better mechanisms for yielding high performance. Bob Beach Ultra Network Technologies ----- End Forwarded Message -----