Socket Migration

[Donald Becker]

On Thu, 16 Aug 2001, Amber Palekar wrote:

>   I , along with my three friends , am planning to
> implement TCP/IP Socket Migration for a distributed
> system ( MOSIX in particular and thats why a kernel
> level implementation of the same ). We have been
> looking for papers for the same .But it seems that not
> much papers are out as yet.

There is a reason why you got detailed responses to the network virtual
memory question, and no responses to migrating TCP sockets.

TCP socket migration is very hard.  It requires fully understanding
TCP/IP and the specifics of the protocol stack implementation.  The
socket state information is distributed across the protocol stack, and
is difficult to gather.  Consider how much code you'll have to touch in
order to just clean up a still-open socket without accidentally opening
the TCP window or responding with a RST to the other end.

Limiting the scope to just a Mosix cluster is significantly easier.  You
have full control over both ends of the connection.  This allows IP
and port address translation at each end, and suspending communication
while a new translation is sychronously established.  But if the
implementation were as easy as that makes it sound, you would see dozens
of papers on the topic.


Donald Becker				becker at scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993