rmmcnulty9 / myptm Goto Github PK

CS 2550 Project Design
Jeff Goldsworthy (jmg199) & Ryan McNulty (rmm81)

Data Structures and Classes: Many different data structures will be used as components
of Java classes. Some of these more notable structures are lock trees with lock coupling,
transaction journals or logs and the Scheduler, DataManager and TransactionManager.

● LockTree - The LockTree will be maintained by the Scheduler. Before granting a lock
   the Scheduler traverses the LockTree putting the appropriate locks (intent, shared,
  or exclusive) at each level until the desired lock is granted.The nodes will have a list
 ordered by timestamp for requested lock and a list for granted locks. The LockTree
will also be used to find deadlocked transactions by comparing the granted locks and
requested locks on each node. We will use Lock Coupling in our Locktree to help
increase concurrency. One down side is our requests will have to go deeper in the tree
to find out if it is blocked. But this time will be made up because fewer locks need to be
released at the end of a transaction.

● Data Files, Pages and Records - Pages will contain Records which come from the
   data files opened by the DataManager. Records will contains the data in each field of
  a given records, the ID, Client Name, and Phone. This structures will be built by the
 DataManager and passed to the Scheduler to become part of the Scheduler’s LockTree.

● Transactions and TransactionManager - The TransactionManager will create
   Transactions by spawning a thread. Using either random or round robin methods of
  concurrent reading and pass the operations into the Transaction for the Scheduler to
 process. Therefore, associated with the Transaction will be its set of operations both
scheduled & unscheduled. The Transaction will also have handles to the locks it holds
and handles to the locks it has requested.

● Scheduler - The Scheduler will manage the LockTree by taking the operations passed
   into it by the TransactionManger through the Transaction. Using the LockTree it will
  obtains locks need for operations putting the appropriate flags in the tree structure and
 return handles for the locks to the Transaction. Once the locks have been acquired the
operations move from being unscheduled to scheduled and the DataManager can then
execute it.

● DataManager and Journals - The DataManager maintains the Pages in the memory
   buffer. It performs either scan or hash search methods on directed files to find the Page
  it needs to load for a given operation. While performing an operation the DataManager
 will store in its Journal (aka log buffer) an undo operation for it. This way in the case of
an abort the DataManager can undo a Transaction’s operation set. The DataManager
 will maintain a buffer management table consisting of the block ID, dirty bit, fix count,
page LSN, Stable-LSN and page number. Additionally the DataManager will use a
modified greedy dual page replacement scheme using the fix count as an indication of
which pages are eligible for replacement in addition to the Least Recently Used and
Least Frequently Used criteria.

Methods: Different schemes will need to be implemented for dealing with aborts and deadlocks.
The following is a discussion of what methods will be used to deal with these problems.

● Undo - Once the DataManager is notified that a transaction has aborted it will use the
   Journal to rollback. Each entry in the Journal will connected as a double linked list using
  the LSN of the previous and next entry. By using checkpoints the DataManager will do
 periodic garbage collection on the Journal, removing committed operations occurring
before the last commit, all aborted or checkpointed operations, and any committed
value matching the value already on stable storage. The checkpointing method used
by the DataManager will be Fuzzy Checkpoints. The system will stop accepting new
operations, then flush to disk only the dirty buffer blocks that were not flushed in the
previous checkpoint and force-write the checkpoint record to the Journal, then finally
resumes normal operation. The trade-off to using this type of checkpoint is that during
a redo the blocked active transactions have to be redone prior to the checkpoint. Since
we are not required to handle redo in this project, this drawback becomes moot.

● Deadlocks - Deadlocks will be detected by creating a deadlock queue containing
   transactions with timestamps of when they last executed an operation. In conjunction
  with this queue a polling timer will be created. When this timer expires it will look at
 the locks granted and requested by any transaction in the deadlock queue that have
expired. For instance, if the timer expires and the transaction T1’s deadlock queue
 timestamp is less than some threshold this means the transaction has not progressed in
the set amount of time and is therefore potentially in deadlock. In this case the locks that
transaction has been granted and has requested are examined. In this examination we
will look for the at least one other transaction that is holding a locks that T1 needs. When
it is found the Wound-Wait algorithm will be used to resolve the deadlock. It is assumed
that undo operations will be very costly which is why we will use Wound-Wait because
the older transaction, presumably with a longer operation history, will never be aborted.

● Strict/Rigorous 2PL - Once the Scheduler receives an abort or commit operations from a
   Transaction by way of the TransactionManager it will access the LockTree and release
  all the locks for that Transaction. This will be the only way locks can be released, by the
 definition of the Strict Industry 2PL scheme.