First Name : Heon
Last Name : Yeom
Title : Associate Professor
Company : Seoul National University
Address : Dept. of Computer Science
City : Seoul
State : Korea
Mail Code : Telephone : +82-2-880-5583
FAX : +82-2-871-4912
E-mail : firstname.lastname@example.org
Cyberpaper : Consistency problem with group commit in distributed main memory database systems. It has been known that the memory resident database, or main memory database (MMDB) has a potential for significant performance advantage over conventional database where the data reside in the disk. The MMDB shows faster response time and improved throughput since there is virtually no I/O overhead for reading and writing the database. The only I/O overhead is from checkpointing and logging for recovery. With the falling price of RAM and the rising demand of high performance system, MMDB's become real alternative for database systems. In most MMDB systems, a small stable memory is used to store the log so that the transactions can commit before the logs are flushed to the disk. When the stable memory is not available, transactions are precommitted.
Pre-committing is done by releasing a transaction's locks as soon as its log record is placed in the log. It helps to improve the throughput since the locks become available without waiting for the log to be flushed to the disk. Along with precommitting, group commit is used to alleviate the log bottleneck. Under group commit, transactions are committed in groups when enough logs are accumulated in the log buffer. Precommit and group commit together can significantly improve the performance of the MMDB by reducing the response time(precommit) and improving the throughput (group commit).
However, when applied to the distributed MMDB, the group commit protocol
combined with the existing two-phased distributed commit protocol may cause
the consistency problem.
To solve this problem, additional round of coordination would be required, which
may greatly degrade the system performance.
We identify the properties of the consistent group commit and
presents a consistent group commit protocol for the distributed main memory