behaviour of innodb_io_capacity contradicts documentation

[aeva black]

Version: percona-server 5.1.56 with innodb 1.0.15-12.5

Description:

According to documentation linked below, and an informal chat with Vadim, there is no need to tune innodb_io_capacity when innodb_adaptive_checkpoint = keep_average. The documentation also suggests that this setting is suitable for SSD storage. In working with several servers with FusionIO storage (single write-master, multiple read-slaves) I have observed that this is not true -- specifically, that changing innodb_io_capacity has a direct impact on flushing of dirty pages and checkpoint age even though innodb_adaptive_checkpoint = keep_average.

http://www.percona.com/doc/percona-server/5.1/scalability/innodb_io.html?id=percona-server:features:innodb_io_51#innodb_adaptive_checkpoint

How to reproduce:

* on replication slave with FusionIO storage, set
   innodb_adaptive_checkpoint = keep_average
   innodb_io_capacity = 100
* then add significant write load to master
* observe innodb checkpoint age stay near max age
* then run mysql> SET GLOBAL INNODB_IO_CAPACITY=10000;
* observe that Innodb_buffer_pool_pages_dirty and checkpoint age both decrease rapidly

Suggested work-around:

Simply tune innodb_io_capacity.

[aeva black (tenbrae)]

I tested this behaviour on the same version of percona-server with innodb_adaptive_checkpoint = estimate, and on hardware with a traditional HDD RAID. In this case, tuning innodb_io_capacity still has an immediate and significant impact on flush rate, which directly contradicts the docs. The docs state that for estimate, "behavior is independent of the innodb_io_capacity variable".

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-2718