Fine-grained Transaction Scheduling in Replicated Databases via Symbolic Execution

Transcription

1 Fine-grained Transaction Scheduling in Replicated Databases via Symbolic Execution Raminhas Stage: 2 nd Year PhD Student Research Area: Dependable and fault-tolerant systems and networks Advisors: Miguel Matos Paolo Romano

2 Research Questions Problem: Replication techniques incur non-negligible costs to maintain consistency among replicas 2PC => Distributed Deadlocks Classic SMR => Serial execution Parallel SMR => Not to trivial to parallelize txs and maintain consistency

3 Research Questions Problem: Replication techniques incur non-negligible costs to maintain consistency among replicas 2PC => Distributed Deadlocks Classic SMR => Serial execution Parallel SMR => Not to trivial to parallelize txs and maintain consistency Why is it a problem: Deadlocks Automatic Conflict Class Prediction => Too Coarse Grained => Level of the table => Low Throughput Manual Prediction => Hard and Not optimal Avoid False negatives => Rollback of Txs, possibly inconsistencies among replicas

4 Research Questions Problem: Replication techniques incur non-negligible costs to maintain consistency among replicas 2PC => Distributed Deadlocks Classic SMR => Serial execution Parallel SMR => Not to trivial to parallelize txs and maintain consistency : Symbolic Execution => Recurring to Symbolic Execution, we are able to provide in a fine-grained and automatic way the set of objects/tuples accessed in a transaction. Why is it a problem: Deadlocks Automatic Conflict Class Prediction => Too Coarse Grained => Level of the table => Low Throughput Manual Prediction => Hard and Not optimal Avoid False negatives => Rollback of Txs, possibly inconsistencies among replicas

5 Research Questions Problem: Replication techniques incur non-negligible costs to maintain consistency among replicas 2PC => Distributed Deadlocks Classic SMR => Serial execution Parallel SMR => Not to trivial to parallelize txs and maintain consistency Why is it a problem: Deadlocks Automatic Conflict Class Prediction => Too Coarse Grained => Level of the table => Low Throughput Manual Prediction => Hard and Not optimal Avoid False negatives => Rollback of Txs, possibly inconsistencies among replicas : Symbolic Execution => Recurring to Symbolic Execution, we are able to provide in a fine-grained and automatic way the set of objects/tuples accessed in a transaction. Consequences: Better Parallelism No Runtime Overhead => offline analysis No False negatives Determinist Scheduling algorithms benefit from fine-grained information

6 Most CPUs are multicore!!! Consensus Some Transactions can be run in Parallel Replicas must remain consistent Schedule must be deterministic C1 C2 C1 C2 C3 C4 C1 C3 C2 C4 C5 C6 C7 C8

7 Compile-Time SE gives the set of keys/tuples accessed Salary Country X Y W S c a n Australia USA Z Tx A: Sum of USA s Sallary : Entire Table s Countries &&

8 Compile-Time SE gives the set of keys/tuples accessed Salary Country X Y W S c a n Australia USA Z Tx A: Sum of USA s Sallary : Entire Table s Countries && Tx B: Increase Sallary :

9 Compile-Time SE gives the set of keys/tuples accessed Salary Country X Y W S c a n Australia USA Z Tx A: Sum of USA s Sallary : Entire Table s Countries && Tx B: Increase Sallary : Disjoint!!!!!

10 Runtime Use fine-grained information to Schedule Transactions Salary Country Tx A Tx B Tx A X Y W S c a n Australia USA Z Tx A: Sum of USA s Sallary : Entire Table s Countries && Tx B: Increase Sallary : Disjoint Lock-Set Run in parallel

11 Runtime Use fine-grained information to Schedule Transactions Salary Country X Y W S c a n Australia USA Z Tx A: Sum of USA s Sallary : Entire Table s Countries && Tx B: Increase Sallary : Solver

12 Lock-based: Solver: + Simple + No runtime overhead - Not trivial to parallelize locktable + Provides the optimal schedule for batch - Imposes runtime overhead

13 Thanks for the attention Q&A