Ad
More Related Content
What's hot (20)
Viewers also liked (9)
Ad
Similar to Mysqlnd query cache plugin benchmark report (20)
Ad
More from Ulf Wendel (20)
Recently uploaded (20)
Mysqlnd query cache plugin benchmark report
- 1. MySQL native driver for PHP: Query cache plugin benchmark impressions
- 2. mysqlnd_qc: Benchmark impressions Ulf Wendel, Andrey Hristov MySQL Connectors Team Sun Microsystems
- 3. Table of Contents Setup and Warning Hardware
- 4. Software Oxid eShop Single machine
- 5. Two machines Artificial tests qc_load.php, mysqlnd_qc_perf.phpt
- 6. Conclusions
- 7. Performance art Performance by number of columns and result set size
- 8. Hardware / Software Box A: Intel Core 2 Duo E6750 @ 2.66 GHz RAID-0, 4 GM RAM
- 9. Apache 2.2.6 (MPM = prefork = default)
- 10. PHP 5.3.3-dev
- 12. APC 3.1.3p1
- 13. Memcache 1.4.3
- 15. Hardware / Software Box B: Athlon 64 X2 Dual Core 5000+ MySQL 5.1.25-rc Source 1 GBit network
- 16. Runs ouf of RAM
- 17. Single machine setup Box A running PHP, Apache, MySQL, Memcache Two machines setup Box A running PHP, Apache, Memcache
- 18. Box B running MySQL
- 19. Oxid eShop 4.2.0 Price winning PHP web shop Community Edition 4.2.0 Generates high CPU and DB load Object-oriented design, many layers
- 21. Smarty
- 22. Demo shop frontpage: ~150 database queries
- 23. Real-life shop frontpages: >>1k database queries
- 24. No practical value? ab -n <x> -c <y> <url>/oxid/index.php Front page only
- 25. No browsing
- 26. No ordering
- 27. Assumption: all queries can be cached Theoretical not practical results Run your own tests!
- 28. We show theoretical maximum, only!
- 29. bench/qc_load.php Run <n> queries for <m> seconds fetch <o> rows per query No practical meaning Storage handler stress testing, only!
- 30. tests/mysqlnd_qc_perf.php Run <n> times SELECT to fetch one row <m> columns
- 31. <o> bytes total per row No practical meaning Storage handler stress testing, only!
- 32. Table of Contents Setup and Warning Hardware
- 33. Software Oxid eShop Single machine
- 34. Two machines Artificial tests qc_load.php, mysqlnd_qc_perf.phpt
- 35. Conclusions
- 36. Oxid, 1 box: PHP/Apache/MySQL CPU bound from 2 clients on, MySQL Server QC wins
- 37. Oxid, 1 box: 1 vs. 10 clients absolut Higher impact once CPU bound, MySQL Server QC wins
- 38. Oxid, 1 box: 1 vs. 10 clients relative Relative: 200% – 400%, MySQL Server QC wins
- 39. MySQLnd QC Plugin vs. Server QC MySQLnd QC Plugin is insecure Can serve stale data
- 40. No knowledge of transactions
- 41. No knowledge of temporary tables
- 42. No knowledge of user variables
- 43. No knowledge of MERGE and stuff
- 45. No protection of cache entries Server QC can cache more For benchmarks set: cache_no_table = 1
- 46. Oxid, 1 box: theoretical maximum Tuned with cache_no_table = 1, PHP QC w. APC wins
- 47. Thoughts on tuning CPU bound from 2 clients on MySQL eats 1 out of 2 cores
- 48. PHP tuning irrelevant, if MySQL eats 1 core
- 49. Cache tuning is irrelevant Goal #1: kick out MySQL! cache all queries!
- 51. cache_no_table = 1 = MySQL Server QC
- 52. Oxid, 1 box: 1 vs. 10 clients tuned Tuned with cache_no_table = 1, PHP QC w. APC wins
- 53. How important is byte code caching? Irrelevant... … if your <Oxid> application is CPU bound
- 54. … and, if you are CPU bound because of MySQL
- 55. … if your <Oxid> application is I/O bound
- 56. … and, if your are I/O bound because of MySQL
- 57. … if But you still want to use it together with QC!
- 58. Oxid, 1 box: PHP byte code caching Byte code caching is nothing without query caching
- 59. Oxid, 1 box: BC gains (1 client) With/without BC compared: 20% - 60% win
- 60. Oxid, 1 box: QC gains (1 client) Without BC, with/without QC compared: 3% - 75% win
- 61. Oxid, 1 box: BC + QC gains (1 client) With BC and QC: 3% - 120% win
- 62. Oxid, 1 box: Use QC and BC together 10 clients: performance improvements up to 180%
- 63. Query caching and your CPU Turn on central heating in the winter
- 64. Turn off fans in summer
- 65. Oxid, 1 box: CPU bound workload Lower MySQL CPU time in favour of PHP CPU time!
- 66. Oxid, 1 box: MySQL CPU usage Share cache contents between PHP worker processes
- 67. Oxid, 1 box: MySQL CPU usage Tuned (cache_no_table=1) to eliminate MySQL usage
- 68. Oxid, 1 box: goal #1 is fewer queries! Monitor MySQLs' Com_select and Queries statistics
- 69. Oxid, 2 boxes: PHP/Apache + MySQL PHP QC and MySQL Server QC about equal fast
- 70. Oxid, 2 boxes: 1 vs. 10 clients Relative: 130 – 200%, CPU bound on PHP/Apache box
- 71. Oxid, 2 boxes: PHP/Apache + MySQL Tuning (cache_no_table=1), theoretical maximum
- 72. Is Memcache storage slow? Invalid benchmark setup! CPU bound: Memcache eats PHP resources
- 73. Why Memcache if using only one PHP machine?
- 74. Needs to be benchmarked with 3+ boxes
- 75. Table of Contents Setup and Warning Hardware
- 76. Software Oxid eShop Single machine
- 77. Two machines Artificial tests qc_load.php, mysqlnd_qc_perf.phpt
- 78. Conclusions
- 79. Performance degration with QC? Rather not - difference is less than 1%
- 80. Which handler is best? Check the basics What is available?
- 81. Share contents among processes?
- 82. Share contents among machines?
- 83. Handler specific features, e.g. statistics All handler seem to scale about equally well … if you consider their general properties
- 84. qc_load.php, 1 box: 20s, 100 rows Significant performance differences
- 85. qc_load.php, 1 box: 20s, 100 rows 1 vs. 10 clients, CPU bound from 2 clients on
- 86. qc_load.php, 1 box: 20s, 100 rows Handler scalability by number of clients
- 87. mysqlnd_qc_perf.phpt, Memcache Average speedup: 1.40
- 88. mysqlnd_qc_perf.phpt, User handler Average speedup: 1.41
- 89. mysqlnd_qc_perf.phpt, APC Average speedup: 1.48
- 90. mysqlnd_qc_perf.phpt, Default Average speedup: 1.55
- 91. Observations CPU bound less MySQL means more PHP
- 92. MySQL % CPU can be reduced by factor 30!
- 93. cache as much as you can!
- 94. cache handler: start with default
- 95. cache handler: step up to APC/Memcache Network bound client cache outperforms server cache
- 96. The End Feedback: [email protected] The End Feedback: [email_address] , [email_address]