8.5.8. Optimizing InnoDB Configuration Variables

Enabling InnoDB to use high-performance memory allocators on systems that include them. See Section 14.2.4.2.12, "Using Operating System Memory Allocators".

Controlling the types of DML operations for which InnoDB buffers the changed data, to avoid frequent small disk writes. See Section 14.2.4.2.13, "Controlling InnoDB Change Buffering". Because the default is to buffer all types of DML operations, only change this setting if you need to reduce the amount of buffering.

Turning the adaptive hash indexing feature on and off. See Section 14.2.4.2.14, "Controlling Adaptive Hash Indexing". You might change this setting during periods of unusual activity, then restore it to its original setting.

Setting a limit on the number of concurrent threads that InnoDB processes, if context switching is a bottleneck. See Section 14.2.4.2.15, "Changes Regarding Thread Concurrency".

Controlling the amount of prefetching that InnoDB does with its read-ahead operations. When the system has unused I/O capacity, more read-ahead can improve the performance of queries. Too much read-ahead can cause periodic drops in performance on a heavily loaded system. See Section 14.2.4.2.16, "Changes in the Read-Ahead Algorithm".

Increasing the number of background threads for read or write operations, if you have a high-end I/O subsystem that is not fully utilized by the default values. See Section 14.2.4.2.17, "Multiple Background InnoDB I/O Threads".

Controlling how much I/O InnoDB performs in the background. See Section 14.2.4.2.20, "Controlling the InnoDB Master Thread I/O Rate". The amount of background I/O is higher than in MySQL 5.1, so you might scale back this setting if you observe periodic drops in performance.

Controlling the algorithm that determines when InnoDB performs certain types of background writes. See Section 14.2.4.2.21, "Controlling the Flushing Rate of Dirty Pages from the InnoDB Buffer Pool". The algorithm works for some types of workloads but not others, so might turn off this setting if you observe periodic drops in performance.

Taking advantage of multicore processors and their cache memory configuration, to minimize delays in context switching. See Section 14.2.4.2.23, "Control of Spin Lock Polling".

Preventing one-time operations such as table scans from interfering with the frequently accessed data stored in the InnoDB buffer cache. See Section 14.2.4.2.24, "Making the Buffer Pool Scan Resistant".

Adjusting your log files to a size that makes sense for reliability and crash recovery. See Section 14.2.4.2.25, "Improvements to Crash Recovery Performance". Historically, people have kept their InnoDB log files small to avoid long startup times after a crash. Internal improvements in InnoDB make startup much faster, so the log file size is not such a performance factor anymore. If your log files are artificially small, increasing the size can help performance by reducing the I/O that occurs as redo log records are recycled.

Configuring the size and number of instances for the InnoDB buffer pool, especially important for systems with multi-gigabyte buffer pools. See Section 14.2.4.2.27, "Improvements to Performance from Multiple Buffer Pools".

Increasing the maximum number of concurrent transactions, which dramatically improves scalability for the busiest databases. See Section 14.2.4.2.28, "Better Scalability with Multiple Rollback Segments". Although this feature does not require any action during day-to-day operation, you must perform a slow shutdown during or after upgrading the database to MySQL 5.5 to enable the higher limit.

Moving purge operations (a type of garbage collection) into a background thread. See Section 14.2.4.2.29, "Better Scalability with Improved Purge Scheduling". To effectively measure the results of this setting, tune the other I/O-related and thread-related configuration settings first.

Reducing the amount of switching that InnoDB does between concurrent threads, so that SQL operations on a busy server do not queue up and form a "traffic jam". Set a value for the innodb_thread_concurrency option, up to approximately 32 for a high-powered modern system. Increase the value for the innodb_concurrency_tickets option, typically to 5000 or so, This combination of options sets a cap on the number of threads that InnoDB processes at any one time, and allows each thread to do substantial work before being swapped out, so that the number of waiting threads stays low and operations can complete without excessive context switching.