C[omp]ute

From: "andrew cooke" <andrew@...>

Date: Thu, 26 Aug 2004 09:30:03 -0400 (CLT)

[Caveat emptor - the summary below is from just 5 minutes scanning these
documents!]

The size of IP packets is not fixed.  The sender can choose the size and
intermediate nodes on the path may fragment the packet.  The receiver must
assemble the fragments correctly.

Since each packet has an associated overhead, it is common to send large
packets.  There is an "experimental" process for finding the maximum
packet length that will avoid fragmentation on any particular route.  It
is described in RFC 1191 - http://www.faqs.org/rfcs/rfc1191.html

However, that process is not always used (it does not always work and is
only a heuristic anyway).  So fragmentation still occurs.  When it does
so, the receiver uses the ID field in the IP packets to re-assemble the
data.  Unfortunately the ID is only 16 bits in size.  At high bandwidths
and data rates this can be insufficient to uniquely identify packets. 
Mistakes are made and communication corrupted.

This is described here -
http://www.ietf.org/internet-drafts/draft-mathis-frag-harmful-00.txt  As
the authors observe, checksums at higher layers should detect and correct
this problem, as should the larger ID field used in IP6.  However, for
"simple", high-speed, high-bandwidth data transfer, this seems like it may
be a serious problem.

From Joseph Lappa on the Global Grid Forum Data Transport mailing list -
http://www.gridforum.org/L_WG/Mailing.htm +
https://forge.gridforum.org/projects/data-rg

Andrew

-- 
` __ _ __ ___  ___| |_____   work web site: http://www.ctio.noao.edu/~andrew
 / _` / _/ _ \/ _ \ / / -_)  personal web site: http://www.acooke.org
 \__,_\__\___/\___/_\_\___|  list: http://www.acooke.org/andrew/compute.html


_______________________________________________
compute mailing list
compute@...
https://200.30.249.238/mailman/listinfo/compute