cck@CUNIXC.COLUMBIA.EDU (Charlie C. Kim) (12/19/87)
When does the fact that KIP doesn't handle IP fragmentation hurt? Steve Fram at CITI suggested trying cross-country (well, not quite cross-country, but it is pretty far from Ann Arbor to Manhattan) Aufs sessions. I found the following. I: o could not speak with their AppleShare server o could write files to, but not read files from their Aufs server The reason was that the KIP boxes do not handle IP fragmentation. The MTU between here and CITI is set at 576 (the arpanet default mtu) or 574 somewhere along the line. The initial startup requires big (602 data bytes) packet to be echoed back and forth. With their AppleShare server it was KIP to KIP and things died. It happened to work with Aufs because the initial packet is sent by the client to a host that handles fragmentation. Writing files to their Aufs server worked because the vax handles IP frags. Reading files from their Aufs server fails because the server uses full size packets and KIP box drops the fragments. Other random failures would also occur (enumerate, GetIcon, etc. can all utilize large packets). Oh well. The main purpose of this note is to warn people who are running between gateways that have MTUs less than 603+sizeof(IP)+sizeof(UDP) that certain session protocols may fail (e.g. ASP). It is not clear to me that the definition of ASP allows one to by-pass this problem. Printing would work fine with MTUs at 576 though since packet sizes are restricted to 512 bytes. Charlie C. Kim User Services Columbia University
jtw@lcs.mit.edu (John Wroclawski) (09/13/89)
In article <37995@bu-cs.BU.EDU> budd@bu-cs.BU.EDU (Philip Budne) writes: From: budd@bu-cs.BU.EDU (Philip Budne) Summary: MacTCP sends bad MSS for localtalk? In article <1989Sep5.121104.10155@gpu.utcs.utoronto.ca> nishri@gpu.utcs.UUCP (Alex Nishri) writes: >The first problem is that the Mac II must be directly on an ethernet >and not on a Localtalk behind a Gatorbox. This is because the >Gatorbox (and Fastpath, we understand) do not handle TCP/IP packet >fragmentation; most Internet SMTP servers use large packet sizes and >these end up on the floor. I seem to remember NCSA Telnet over localtalk sends a TCP MSS option requesting minimal (576) maximum segment size. The one MacTCP application I looked at sends an MSS request of 1024 when running over localtalk. Actually, Gatorboxes are supposed to fragment IP packets correctly. (Please let Cayman know if you think otherwise...). Most older Mac IP implementations don't do reassembly, though. IMHO it would still be better if MacTCP requested a small enough MSS to avoid fragmentation. Aside from not working with some current Ethernet-Localtalk gateways, I suspect that some host hardware can't handle the back-to-back packets and will often drop the second fragment.