Mark Callaghan at facebook tested the test release of MySQL 5.6.3 and he has found some performance improvement with InoDB feature. Read below...
Mark tried two of the previews for MySQL 5.6.3 at http://labs.mysql.com/. His first attempt with the multi-threaded slave preview was not successful. Parallel apply on the slave is serial when the master does not also run 5.6.3. He said (I hope this changes as a typical upgrade is first done on the slave.)
Mark tried two of the previews for MySQL 5.6.3 at http://labs.mysql.com/. His first attempt with the multi-threaded slave preview was not successful. Parallel apply on the slave is serial when the master does not also run 5.6.3. He said (I hope this changes as a typical upgrade is first done on the slave.)
He was more successful with the InnoDB features preview. A few more mutex contention bottlenecks were removed in it and he wanted to compare the peak row update rate between it and MySQL 5.1.52. he configured InnoDB to use a buffer pool large enough to cache all data and ran a custom version of sysbench with 8 tables. The peak rate on the preview is about twice the peak rate on the unmodified InnoDB plugin in 5.1.52 using an 8-core server.
This is good news. The results below list the number of rows updated per second using 8 to 256 concurrent clients updating 1 row by primary key per UPDATE statement.
Configuration used:
The database had 8 tables with 2M rows each.
The binlog was disabled during the test.
innodb_flush_log_at_trx_commit=2,
innodb_doublewrite=0.
This is a configuration meant for benchmarks but it also allows maximum stress to be put on InnoDB. He only ran the test once for each level of concurrency and won't try to explain the results at 32 connections.
8 | 16 | 32 | 64 | 128 | 256 | Clients |
32423 | 60394 | 121480 | 77951 | 76189 | 74127 | mysql 5.1.52 |
30446 | 53512 | 107291 | 138408 | 149003 | 151705 | mysql 5.6.3 |
0 comments:
Post a Comment