| Bug #72466 | More memory overhead per page in the InnoDB buffer pool | ||
|---|---|---|---|
| Submitted: | 26 Apr 2014 14:57 | Modified: | 28 Apr 2014 6:23 |
| Reporter: | Mark Callaghan | Email Updates: | |
| Status: | Verified | Impact on me: | |
| Category: | MySQL Server: InnoDB storage engine | Severity: | S3 (Non-critical) |
| Version: | 5.7.4 | OS: | Any |
| Assigned to: | CPU Architecture: | Any | |
[26 Apr 2014 15:04]
Mark Callaghan
5.6.12 includes the Facebook patch. Asking others whether we did anything to use less memory
[26 Apr 2014 15:11]
Mark Callaghan
This is something we made better in 5.6 https://github.com/facebook/mysql-5.6/commit/91e979e8436b83400e918fa0f251036e50d0cb5f
[26 Apr 2014 15:12]
Mark Callaghan
So this is a feature request, not a bug report.
[28 Apr 2014 6:22]
MySQL Verification Team
Hello Mark, Thank you for the feature request! Thanks, Umesh
[28 Apr 2014 6:42]
Sunny Bains
For v5.7 you can use -DMUTEX_TYPE="futex" to get rid of the event in a mutex on Linux systems.
[28 Apr 2014 7:17]
Sunny Bains
There is also this unused member variable in os0event.cc.
=== modified file 'storage/innobase/os/os0event.cc'
--- storage/innobase/os/os0event.cc revid:tor.didriksen@oracle.com-20140423150641-94y77lkeyak7vg3f
+++ storage/innobase/os/os0event.cc 2014-04-28 07:23:15 +0000
@@ -84,7 +84,6 @@ typedef pthread_cond_t os_cond_t;
#endif /* _WIN32 */
typedef std::list<os_event_t> os_event_list_t;
-typedef os_event_list_t::iterator event_iter_t;
/** InnoDB condition variable. */
struct os_event {
@@ -289,10 +288,6 @@ private:
os_cond_t cond_var; /*!< condition variable is
used in waiting for the event */
-
-public:
- event_iter_t event_iter; /*!< For O(1) removal from
- list */
protected:
// Disable copying
os_event(const os_event&);
Description: Looking at memory overhead for InnoDB with innodb_page_size=4kb between 5.6.12 and 5.7.4. Used the same build options (PS disabled) and same my.cnf with one exception. * 5.6 - buffer_pool_size=110G * 5.7- buffer_pool_size=105G Looking at VSZ and RSS from ps output shows that 5.7 uses about 2.7% more memory for 5G less in the buffer pool. Normalizing for the differences in buffer_pool_size it uses 7.5% more memory for the same size buffer pool. VSZ RSS 5.6 -> 132272564 131386384 5.7 -> 135881780 134215368 How to repeat: my.cnf for 5.7.4 listed below sql_mode=NO_ENGINE_SUBSTITUTION,STRICT_TRANS_TABLES table-definition-cache=1000 table-open-cache=2000 table-open-cache-instances=1 max_connections=2000 key_buffer_size=200M metadata_locks_hash_instances=256 query_cache_size=0 query_cache_type=0 skip_log_bin innodb_buffer_pool_instances=8 innodb_io_capacity=16000 innodb_lru_scan_depth=2000 innodb_checksum_algorithm=CRC32 innodb_thread_concurrency=0 # innodb_log_file_size=1900M innodb_log_file_size=3900M innodb_flush_log_at_trx_commit=2 innodb_doublewrite=0 innodb_thread_concurrency=0 innodb_max_dirty_pages_pct=80 innodb_max_dirty_pages_pct_lwm=0 innodb_file_format=barracuda innodb_file_per_table default_tmp_storage_engine=MyISAM innodb_purge_threads=8 innodb_compression_pad_pct_max=37 innodb_compression_failure_threshold_pct=10 innodb_old_blocks_pct=50 innodb_old_blocks_time=1000 innodb_adaptive_flushing=1 innodb_adaptive_flushing_lwm=10 innodb_flushing_avg_loops=30 eq_range_index_dive_limit=1000 lock_wait_timeout=60 innodb_compression_level=1 innodb_lock_wait_timeout=2 innodb_log_buffer_size=4M innodb_log_compressed_pages=0 innodb_read_io_threads=8 innodb_write_io_threads=8 innodb_stats_on_metadata=0 innodb_page_size=4k innodb_doublewrite=0 innodb_flush_log_at_trx_commit=2 innodb_flush_neighbors=0 innodb_flush_method=O_DIRECT innodb_buffer_pool_size=105G Suggested fix: use less memory