Projects > Concurrency Control Schemes

filterKey: project

H-Store gets part of its performance boost by avoiding the fine-grained concurrency control used by traditional databases.  When data is stored in main-memory, the cost of acquiring and releasing locks is similar to the cost of accessing data. Instead of locks, H-Store simply executes one transaction at a time, from start to finish, without blocking. In order to use multiple machines or CPUs, the database is partitioned into independent pieces, so each partition can execute transactions independently.

This works until there is a transaction that must access data in multiple partitions. H-Store is designed for applications where this is rare, but it still happens. In this case, a partition must stop performing work in order to wait for data from other partitions. This “pause” can kill performance. We have investigated techniques for avoiding this stall, and concluded that using speculation can improve performance over traditional locking based approaches.

Downloads

Low Overhead Concurrency Control: SIGMOD 2010 Experiment Repeatability Code: The source code and the directions for repeating the experiments in the SIGMOD 2010 paper.

Publications:

• E. P. C. Jones, D. J. Abadi, and S. Madden, "Low Overhead Concurrency Control for Partitioned Main Memory Databases," in SIGMOD ’10: Proceedings of the 2010 international conference on Management of data, New York, NY, USA, 2010, pp. 603-614. [PDF] [BIBTEX]
  @inproceedings{jones2010,
  author = {Jones, Evan P.C. and Abadi, Daniel J. and Madden, Samuel},
  title = {Low Overhead Concurrency Control for Partitioned Main Memory Databases},
  booktitle = {SIGMOD '10: Proceedings of the 2010 international conference on Management of data},
  year = {2010},
  isbn = {978-1-4503-0032-2},
  pages = {603--614},
  location = {Indianapolis, Indiana, USA},
  doi = {http://doi.acm.org/10.1145/1807167.1807233},
  publisher = {ACM},
  address = {New York, NY, USA},
  url = {/papers/hstore-cc.pdf},
 }