Messages in this thread | | | Subject | Re: Is something wrong with 2.0.33?? | From | Mark Gray <> | Date | 09 Jan 1998 05:52:13 -0500 |
| |
> I wrote: > > 2. I have watched a tcpdump during other transfers and have found > > that my ACK's after having lost a packet never seem to make it to the > > site I am FTPing from. Doing a traceroute to the site at this time > > will often kick the ftp going again (and I have a little shell script > > I use which watches the file I am getting and does a traceroute should > > it ever stall.) Idle speculation: the route back to the site has > > gotten clobbered and the ICMP package has a better chance discovering > > a new route than does a TCP ack. Now I constantly run this in a very > alan@lxorguk.ukuu.org.uk (Alan Cox) writes: > Idle but wildly implausible. It may just be the time you see drop > outs.
This might be it then (bare with me) --- given that Lucent claims that my aging USR33.6 modem has a compression bug that they are now compensating for on their end, it is possible that sometimes the kernel has been sending back its ACK at a point in time when the two modems compression tables are in a bad state and the packet is (was) either getting lost in the modems or is mangled enough that the packet is dropped the next time someone does a checksum on it. Perhaps the kernel is ever so slightly faster (or slower) turning around when it is ACKing an earlier packet because of the missing one and this could be a timing reason for why it is catching the modems at an inopportune moment. Running the traceroute would then act to modify the compression tables and the next ACK can get through.
This same problem happened to me several times during the "1.3.almost-ready" kernels. At that time my ISP was entirely an USR33.6 shop, so Lucents claims might have a touch of weight (it is always somewhat hard to swallow finger pointing when a change on the other side seems to fix things :-) I have fallen off the list several times since then, so I do not really know whether Linux successfully addressed this problem since then or whether it kept disappearing mysteriously as it used to do during 1.3.* days.
(I will shut up now and finish reading Stevens) [snip] > > Alan > >
| |