RE: Informal Testing of Microsoft's TCPLW & SACK Implementations

From: Eric Travis ([email protected])
Date: Mon Feb 09 1998 - 14:38:17 EST


Jim,

>After reading your trials in changing the TCP window size in Windows98
>beta, the Internet protocols team at NASA Lewis Research Center decided
>to have a go at it. Using two machines in our TCP/HTTP/Multicast
>testbed, we verified Windows98 is capable of using large windows and
>SACKs.

Well, please don't take this the wrong way (it's great that you were
able to do this) but all that it illustrates is that Windows98 Beta
can *advertise* large windows and SACK. Necessary but not sufficient
to make any further judgements about capabilities. Writing a check
doesn't necessarily mean you have the means of covering the funds. :o)

For example, we've seen a problem on at least one RFC-1323 implementation
(FreeBSD-2.2 don't know yet about later releases) where a sufficiently
large buffer space was requested, receive windows advertised correctly
but nevertheless, the OS was not allocating the necessary memory to the
sending socket buffer to prevent the connections from becoming send-window
limited. Chances are that the receiver had the same problem, but there
was never enough data in-flight to verify it. The behavior was persistent
and recreateable. This was on a 2Mbps connection with a .5 sec round trip,
far less strenuous than Steve's scenario.

We haven't determined *exactly* what is causing this incorrect behavior,
(the kernels were configured correctly and the machines had more than
adequate resources) but this is a case where an implementation *is*
RFC-1323 capable, but OS was not providing the necessary resources.

Bugs happen even in mature code and the RFC-1323 functionality is probably
amongst the least tested/exercised portion of most TCP implementations.

If you or Steve could provide tcpdump logs of a representative run (say,
4Mbps pipe with the 1.3s delay) with Windows98 both throwing *and*
catching or one run of each we could then get a better idea of whether or
not things are working properly. Making the logs available for ftp access
is preferable to posting on the list or e-mailing.

I'd do the runs myself but I don't have access to a copy of Windows98 Beta
(delays and errors are trivial in comparison).

Eric



This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:35 EST