Path: utzoo!mnetor!uunet!husc6!mailrus!ames!pasteur!ucbvax!VAX.FTP.COM!jbvb
From: jbvb@VAX.FTP.COM (James Van Bokkelen)
Newsgroups: comp.protocols.tcp-ip.ibmpc
Subject: Re: Pc/Ip scorecard - part II
Message-ID: <8803021414.AA26951@vax.ftp.com>
Date: 2 Mar 88 14:14:43 GMT
References: <8803020725.aa20913@Louie.UDEL.EDU>
Sender: usenet@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 39

Reassembly doesn't help you much unless you have an Ethernet board with
multiple packet buffers and relatively fast logic.  We added reassembly
support (only two fragments, resulting packet must be smaller than the
network MTU) for someone who had a Bridge GS-7? that fragmented everything
bigger than 500 bytes of IP length.  The result works, and it is in the
code we ship, but it was just about useless with the 3C500s he had.

The problem is that gateway developers fragment the packet immediately
before sending it, and they pride themselves on how fast they make their
Ethernet interfaces go.  You can't get a single-buffered interface ready
in time for the 2nd fragment, period.  Pairs of fragments are very good
ways of determining just how good your multiple-buffered interface is.

I don't have high-precision instrumentation to measure intervals, but
I don't believe that any of the PC interfaces available today can grab
two packets as fast as some other vendors send them.  I have also heard
criticism (possibly inaccurate) of some Unix vendors for sending packets
closer together than the Ethernet spec allows.

I think generating fragments from a PC is a bad idea, and should be avoided.
Unless you can arrange to send a re-transmission with the same IP ID field,
sending fragments doubles the dropped-packet rate once for each fragmentation.
We deal with this issue by requesting a smaller TCP MSS when a connection
is routed off the local subnet.  Our TCP window is also settable.  Philip
Prindeville's version of the CMU code would seem to have to generate
fragments, because he supports TFTP on ARCnet, with its 508-byte MTU, but
I haven't looked at it or used it.

Speaking of subnets, I think you should have a "supports subnets" column.
Our code does, and I know that Excelan, at least, doesn't.  MIT-derived code
ought to support them, vendors with older on-board TCPs are less likely to.
I also think that 'subnets' settable on 8-bit boundaries only is not "subnet
support".

	    IP frag Subnet
FTP	   | reass | yes  |.....

James B. VanBokkelen
FTP Software Inc.