Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!iuvax!rutgers!elbereth.rutgers.edu!hardees.rutgers.edu!patterso From: patterso@hardees.rutgers.edu (Ross Patterson) Newsgroups: comp.protocols.tcp-ip Subject: Re: Reliability of TCP/IP Message-ID: Date: 30 Aug 89 17:25:00 GMT References: <8908281553.AA01532@interlan.interlan.com> Organization: Rutgers Univ., CCIS Lines: 21 Frank Kastenholz writes: >TCP/IP as a suite IS reliable (or at least as reliable as the rest of the >network world:-). Where the confusion may have arisen is that IP, in and >of itself is NOT reliable. Reliable has several definitions. Webster offers "suitable or fit to be relied on" and "giving the same result on successive trials". TCP as defined in the RFCs meets both, but as implemented in many cases, fails the latter. In particular, the "keepalive abomination" causes connections that might otherwise survive brief outages or reconfigurations. MIS departments (the original message cited one) generally understand a third definition, embodied in the IBMism "RAS" (Reliability, Availability and Servicability): "safe, tested, designed to avoid problems". Any system that aborts a 100MByte FTP that's already 85% complete because an intervening gateway was rebooted flunks this. Proper use of exponential backoff and avoidance of keepalives put the R back in RAS. Just ask Phil ("I Hate KeepAlives") Karns, who has reported FTP sessions that succeeded after being interrupted for *days* in the ham radio world, where such things are done right. Ross Patterson Rutgers University