Rocks and Racks
Victor C. Zandy   (zandy@cs.wisc.edu)
Barton P. Miller  (bart@cs.wisc.edu)
University of Wisconsin-Madison, USA


We shall demonstrate two new systems: reliable sockets (rocks)
and reliable packets (racks).  Both systems provide transparent
network connection mobility to ordinary applications using only
user-level mechanisms.  These systems have applications in
end-user mobile computing and process migration of parallel
programs.


Their major functionality is:


1 They detect a connection failure within seconds of its
  occurrence;


2 They preserve the endpoints of a failed connection in a
  suspended state for an arbitrary period of time;


3 They automatically reconnect as connectivity is restored, even
  when one end changes IP address, with correct recovery of
  in-flight data;


4 They safely interoperate with ordinary clients and servers,
  using a new user-level Enhancement Detection Protocol.


Rocks and racks have different implementation models.  Rocks
intercept and modify the behavior of the sockets API with a
library that is interposed between the application and the
operating systems.  Racks use a packet filter controlled by a
separate daemon to intercept and modify the packets exchanged
over a connection.  Both techniques enable transparent network
connection mobility without any kernel modifications.


We will demonstrate the mobile computing applications of rocks
and racks, particularly the ability of remote shells and remote X
Windows applications to survive periods of disconnection and
reconnection from new addresses.


We also will demonstrate how rocks and racks support process
migration of programs that communicate over sockets.  In
particular, we have extended a basic process checkpointing
mechanism to checkpoint parallel MPI programs and to enable the
Condor high throughput computing environment to migrate jobs with
open network connections.

More information about rocks and racks is available at
http://www.cs.wisc.edu/~zandy/rocks