Distributed Transaction Processing in the Escada Protocol

Transcription

1 Distributed Transaction Processing in the Escada Protocol Alfrânio T. Correia Júnior Grupo de Sistemas Distribuídos Departamento de Informática Universidade do Minho Distributed Transaction Processing in theescada Protocol p.1/33

2 Scenario Dependable Databases are core components of our Information Society. Software based replication is an attractive solution to assure dependability. Problems arise when attempting to preserve the following: Strong consistency (1SR). Updates are carried through any replica. Distributed Transaction Processing in theescada Protocol p.2/33

3 Scenario (Problems) Traditional database replication protocols do not scale up well due to: The high number of messages exchanged among the replicas. The deadlock rate proportional to n 3, where n is the number of replicas, which is impractical. Distributed Transaction Processing in theescada Protocol p.3/33

4 Scenario (Earlier Solutions) Lazy Replication relaxes the consistency criteria. Master/Slave chooses the replica that can receive the updates. Distributed Transaction Processing in theescada Protocol p.4/33

5 Escada Approach Pattern on the DBSM Replication based on group communication Update-everywhere and deferred updates Atomic broadcast to propagate the transaction s processing data Total order is combined with a conflict detection process to assure 1SR. Distributed Transaction Processing in theescada Protocol p.5/33

6 Protocol Overview S1 S2 T1 02 T Local Execution 02 - Atomic Multicast ( RS, WS, WV ) 03 - Certification Termination Protocol Distributed Transaction Processing in theescada Protocol p.6/33

7 Protocol Overview A, B, C, D, E A, B, C, D, E S1 S2 T1 S3 A, B, C, D, E T1 = { A, B, C } T2 = { C, D, E } T2 Distributed Transaction Processing in theescada Protocol p.7/33

8 Escada Approach Target: Large scale distributed systems. Provides partial replication (Partial DBSM) Reduces resource consumption exploiting application s locality. Distributed Transaction Processing in theescada Protocol p.8/33

9 Motivation How the Escada can be augmented to provide partial replication? Distributed Transaction Processing in theescada Protocol p.9/33

10 Main Contributions Distributed Transaction Processing Semantic Caching Extensions to the PostgreSQL Evaluation Process (TPC-C and TPC-W) Distributed Transaction Processing in theescada Protocol p.10/33

11 Contribution Distributed Transaction Processing Distributed Transaction Processing in theescada Protocol p.11/33

12 Partial DBSM Execution: A site that handles a transaction may not be able to locally complete its execution. It is possible that no single site can do it. Termination Protocol: The distributed execution fragments the knowledge about conflicts. How to decide if a transaction can commit or not? Distributed Transaction Processing in theescada Protocol p.12/33

13 Execution Rewrites the queries mapping the original relations to the actual fragments. Sites are contacted according to the replicated fragments. Distributed Transaction Processing in theescada Protocol p.13/33

14 Termination Protocol The updated information is propagated to the pertinent replicas. What can we say about the read and write sets? They could be sent to all the replicas allowing a deterministic certification (non-voting). They could be sent just to the pertinent replicas requiring a voting certification. Distributed Transaction Processing in theescada Protocol p.14/33

15 Non-Voting A, B, C, D, E A, B, C, D, E S1 S2 T1 S3 A, B, C, D, E T1 = { A, B, C } T2 = { C, D, E } T2 Distributed Transaction Processing in theescada Protocol p.15/33

16 Voting A B S1 S2 T1 C,D,E S3 T1 = { A, B, C } T2 = { C, D, E } T2 Distributed Transaction Processing in theescada Protocol p.16/33

17 Experiments Latency Latency (ms) DBSM - WAN PDBSM - WAN PDBSMRAC - WAN DDE - WAN Ideal Locking Clients Distributed Transaction Processing in theescada Protocol p.17/33

18 Contribution Semantic Caching Distributed Transaction Processing in theescada Protocol p.18/33

19 Semantic Caching Aims at reducing communication among distributed replicas. Exploits the broadcast to build the refresh mechanisms. Reduces the management overhead when compared to tuple and page-based solutions. Roughly, it caches the result sets and identifies them based on predicates. Distributed Transaction Processing in theescada Protocol p.19/33

20 Semantic Caching The satisfiability background is behind this approach. Two predicates S and F are satisfiable if S F is not a contradiction. The main idea: "upon receiving a query is possible to use a previous request?". Distributed Transaction Processing in theescada Protocol p.20/33

21 Semantic Caching We use an important class of queries, called SPJ queries. However, other queries could be cached. Distributed Transaction Processing in theescada Protocol p.21/33

22 Experiments Time Elapsed (ms) Profits 5 Clients 10 Clients 15 Clients 20 Clients Cache Entries (x100) Distributed Transaction Processing in theescada Protocol p.22/33

23 Contribution PostgreSQL Distributed Transaction Processing in theescada Protocol p.23/33

24 PostgreSQL We identified that the 1SR correctness criteria suggested by the DBSM is not achievable. We present an informal algorithm about how to exactly extract the read sets. Finally, we show how to extend the "rule mechanisms" of the PostgreSQL. Distributed Transaction Processing in theescada Protocol p.24/33

25 Contribution Evaluation Process Distributed Transaction Processing in theescada Protocol p.25/33

26 Evaluation Process We use a simulation tool that allows the combination of real and simulated code. The replication protocols, our main goal, are real implementations. This approach permits us to focus on our goal while at the same allows: Variations on the network architecture, the workload, concurrency control policies. Distributed Transaction Processing in theescada Protocol p.26/33

27 Evaluation Process TPC-W mimics an Internet commerce application. TPC-C mimics a wholesale supplier (OLTP). Distributed Transaction Processing in theescada Protocol p.27/33

28 DBSM Latency DBSM & Centralized Latency (ms) DBSM - LAN 1 CPU 3 CPUs Clients Distributed Transaction Processing in theescada Protocol p.28/33

29 Partial DBSM Latency Latency (ms) DBSM - WAN PDBSM - WAN PDBSMRAC - WAN DDE - WAN Ideal Locking Clients Distributed Transaction Processing in theescada Protocol p.29/33

30 Partial DBSM Kbytes/s Network Bandwidth DBSM - WAN PDBSM - WAN Clients Distributed Transaction Processing in theescada Protocol p.30/33

31 Conclusion Database replication based on group communication is a feasible solution. The non-voting solution scales up better. The semantic caching reduces communication. Partial replication reduces resource consumption. Distributed Transaction Processing in theescada Protocol p.31/33

32 Future Work Improve our simulation tool: user friendly. Develop a version of our semantic caching to mobile elements. Revisit the Epsilon Serializability to Improve DBSM performance. Distributed Transaction Processing in theescada Protocol p.32/33

33 The End Unfortunately, I don t know how to the things that I need to do. However, as a good start point I certainly know how I must not. Sometimes, this knowledge avoids worthless efforts. Sometimes, we simply need to go into the wrong direction to choose another path. Distributed Transaction Processing in theescada Protocol p.33/33