How to enforce correct runtime execution?

[rickyyx]

For each transaction worker:

• init with a list of pieces and the conflict graph
• get the next runnable piece
• run the piece (potential race?)

Implementation details:

How does a worker get the workload?

• The chopper(test chopper) should register the workload
• Each worker will then get a list of transactions consisting of pieces
• Each worker will also get part of the conflict graph.

How to find out if a piece is runnable?

• Global data structure keeps track of what transaction programs are being run by which transaction instances and transaction worker. (Tid)
• Check the progress of that transaction worker, ($
  • by checking the log? (presence of log means the piece has already started)
  • by checking some bits?
• Cannot run if there are existing dependencies conflicting pieces($dep list) that have not run => delay

When running piece, what to update?

Dependency list:

• In the previous step, whenever found a transaction instance that has already ran a conflicting piece, add it into the $dep list

Progress :

• this has to be in sync with the progress checking.

TODO:

• [x] Implement hard-coded chopped pieces and conflict graph.
• [x] Implement transaction worker.
• [x] multiple txn instances of the same transaction base need to handle conflicts of itself.
• [ ] insertions, how to lock it and do persistence?

[rickyyx]

How to do chopping?

Option 1: (for code release?)

• Wrap statement (if-else, read, write, loops) and construct graphs.

Option 2: (for own benchmark)

• Manually generate pieces (no chopping is needed) and dependency graph

[rickyyx]

How to do logging?

Option 1: Logging all writes at the beginning of the transaction and wrap them into a piece.

• Why this doesn't work?: Each log will require a read, and thus this read will potentially have a conflict edge with a write in another transaction. The additional conflict edge will expand the strongly connected graph, and previously independent pieces now have to be merged into a mega piece...

Option 2: Logging at the beginning of each piece.

• This might incur additional delay due to fencing after each logs.