1. H-Store: A Specialized Architecture for High-throughput OLTP Applications Evan Jones (MIT) Andrew Pavlo (Brown) 13 th Intl. Workshop on High Performance Transaction Systems October 26, 2009

2. Intel Xeon E5540 (Nehalem/Core i7) October 26, 2009 Source: Intel 64 and IA-32 Architectures Optimization Reference Manual

3. Distributed Clusters October 26, 2009

4. Scaling OLTP on Multi-Core? Use a distributed shared-nothing design October 26, 2009

5. How to Make a Faster OLTP DBMS  Main memory storage  Replication for durability  Explicitly partition the data  Specialized concurrency control  Stored procedures only  Single partition: execute one transaction at time  Multiple partitions: supported but slow October 26, 2009

6. Source: S. Harizopoulos, D. J. Abadi, S. Madden, M. Stonebraker, “OLTP Under the Looking Glass”, SIGMOD 2008. OLTP: Where does the time go? October 26, 2009

7. Users Rely on Partitioning October 26, 2009 Source: R. Shoup, D. Pritchett, “The eBay Architecture,” SD Forum, Nov. 2006.

8. What about multi-core?  Traditional approach:  One database process  Thread per connection  Shared-memory, locks and latches  H-Store approach:  Thread per partition  Distributed transactions October 26, 2009

9. Example Microbenchmark  One table per client Table(id INTEGER, counter INTEGER)  Each client executes the following query: UPDATE Table SET counter = counter + 1 WHERE id = 0;  Add clients to find maximum throughput  Data on RAM disk October 26, 2009

10. Experimental Configuration October 26, 2009 ThreadsPartitions

11. Partitions versus Threads October 26, 2009 Ideal Partitions Threads

12. Scalability Analysis Partitions scale better than threads.  Threads: contention for shared resources [1]  Partitions: memory bottleneck causes sublinear scaling H-Store: Not just for distributed shared-nothing clusters October 26, 2009 [1] R. Johnson et al., "Shore-MT: A Scalable Storage Manager for the Multicore Era," EDBT 2009.

13. Multi-core Design Problem  How to automatically create a data placement scheme to improve multi-core throughput?  Data Partitioning:  Maximize the number of single-partition transactions.  Data Placement:  Maximize the number of single-node transactions. October 26, 2009

14. Database Partitioning  Select partitioning keys and construct schema tree. October 26, 2009 DISTRICT CUSTOMER ORDER_ITEM ITEM STOCK WAREHOUSE ORDERS DISTRICT CUSTOMER ORDER_ITEM STOCK ORDERS ITEM Replicated WAREHOUSE TPC-C Schema Schema Tree

15. ITEM ITEMj ITEM Database Partitioning  Combine table fragments into partitions. October 26, 2009 P2P4 DISTRICT CUSTOMER ORDER_ITEM STOCK ORDERS Replicated WAREHOUSE P1 P2 P3 P4 P5 P3P1 ITEM Partitions ITEM Schema Tree

16. Partition Affinity Core1Core2 HT2 HT1 Core1Core2 HT2 HT1 Data Placement  Assign partitions to cores on each node. October 26, 2009 P1 ITEM P2 ITEM P5 ITEM Partitions Cluster Node P4 ITEM P3 ITEM Node 1Node n P1 P2 P3 P4 P5

17. Core1Core2 HT2 HT1 Core1Core2 HT2 HT1 Data Placement  Assign partitions to cores on each node. October 26, 2009 P1 ITEM P2 ITEM P5 ITEM Partitions Cluster Node P4 ITEM P3 ITEM Node 1Node n P1 P2 P3 P4 P5 Partition Affinity P1+P5 ITEM

18. H-Store’s Future October 26, 2009

19. Fast Databases October 26, 2009 Larry Ellison

20. H-Store’s Future  New Name. New Company.  Six full-time developers.  Open-source project (GPL)  Beta by end of 2009  Multiple deployments in financial service areas. October 26, 2009

21. More Information  H-Store Info + Papers:  http://db.cs.yale.edu/hstore/  VoltDB Project Information:  http://www.voltdb.org/ October 26, 2009