19 May 2003

Progress Report: Fault Tolerant Matlab*P 

- Nisheeth Shrivastava
- Rachit Chawla

-----------------------------------------------
The project consists of the following features:
-----------------------------------------------    
    * A shadow Matlab*P server running on a fault free machine. 
	- It will work as a mediator between client and MatlabXP server. 
	
	-----------------------
	The main tasks will be:
	-----------------------
	- Open client (Matlab) connection.
	- Start MatlabXP servers.
	- Take periodic checkpoint.
	- Recover the servers in case of any fault.
    
    * Checkpointing library integrated into MatlabXP code.
        - Take checkpoint on "CKPT" message from shadow server.
	- Recover if Shadow server detects any process failure.

-----------
Tasks Done:
-----------

-- Shadow Server Ready
----------------------
   -------
   Part 1:
   -------
   -- Matlab Client connects to Shadow Server
   -- Shadow start MatlabXP as one of its child process
   -- Shadow Server connects to MatlabXP
   -- Matlab Client sends command to Shadow Server
   -- Shadow Server send command to MatlabXP Server
   -- Shadow gets the results and sends it to MatlabXP
   
   -------
   Part 2:
   -------
   -- Shadow checkpoints MatlabXP server, (currently) when an external 
signal 
      is sent.  
   -- If it detects any failure, restarts the MatlabXP server from the
      last checkpoint.
   -- Opens all connections again with new MatlabXP server, to keep the
      recovery transparent to the matlab client.
   -- Send the last command after failure again, get the results and send
      it to the matlab client.

-- Checkpoint library
---------------------
	We have implemented a checkpointing library for Linux system.
	It has been tested to successfully checkpoint and recover a normal 
	process. We are currently on the stage to make it work for any MPI 
	process.

   -- Checkpointing can be done for a MPI Process.
   -- Recovery part in progress.

----------------
Tasks to be done:
-----------------

-- Recovering a MPI process from the saved checkpoint.
-- Checkpointing MatlabXP server periodically w/o any external signal, on
   a "CKPT" message from Shadow Server.
-- Sending the last command after failure again, get the results and send
   it to the matlab client.
-- Further Testing.