Distributed Transaction Management, Autumn 2003 , Exercise 3
Due Tue 11.11.2003

1. Change the LockManager implementation so that a client first
   contacts one server, which gives it a txn id. Then, when the
   client opens a connection to the other servers, it first
   tells them its txn id.

   One possible way to form a txn id of type Long goes as follows:
   Suppose the servers are given short integer ids. Take the
   current time as Long, shift it left by 16 bits, and put the
   server id in the Long as the 16 rightmost bits. Then you get
   ids, which can be used to order txns by age and ties are
   solved by using the server id.

   LockManager implementation can be found from 
   http://www.cs.uta.fi/~jyrki/dtm03/LockManager.zip

2. Improve the implementation so that all servers know all other
   servers ids (e.g. by reading them from a configuration file, say,
   servers.txt). Then implement deadlock detection by using the
   method whereby the other servers just send their lock information
   to one of the servers, which constructs the overall lock graph.
   For testing purposes, you may just make one of the clients initiate
   the deadlock detection.

3. Consider decentralised 2PC. Suppose that when the txn is ready, someone
   (e.g. a client) initiates the voting by sending a Vote-Request to
   all servers (participants). When would using a termination protocol be
   appropriate and what should it be like?

4. Consider the example LockManager implementation. (a) Suppose that a 2PC
   implementation is added there and locks are released implicitly at
   commit/rollback. When would it make sense for a server
   to vote for No for a transaction (now that we are only managing locks).
   (b) The same question assuming that locks should be released
   explicitly before commit/rollback.

5. Add basic 2PC to your implementation (you may leave out termination
   and recovery and log management for now).