💎一站式轻松地调用各大LLM模型接口,支持GPT4、智谱、星火、月之暗面及文生图 广告
## 代价模型 mysql 5.7代价计算相对之前的版本有较大的改进。例如 * 代价模型参数可以动态配置,可以适应不同的硬件 * 区分考虑数据在内存和在磁盘中的代价 * 代价精度提升为浮点型 * jion计算时不仅要考虑condition,还要考虑condition上的filter,具体参见参数condition_fanout_filter 5.7 在代价类型上分为io,cpu和memory, 5.7的代价模型还在完善中,memory的代价虽然已经收集了,但还没有没有计算在最终的代价中。 5.7 在源码上对代价模型进行了大量重构,代价分为server层和engine层。server层主要是cpu的代价,而engine层主要是io的代价。 5.7 引入了两个系统表mysql.server_cost和mysql.engine_cost来分别配置这两个层的代价。 以下分析均基于mysql5.7.10 ## server_cost * row_evaluate_cost (default 0.2) 计算符合条件的行的代价,行数越多,此项代价越大 * memory_temptable_create_cost (default 2.0) 内存临时表的创建代价 * memory_temptable_row_cost (default 0.2) 内存临时表的行代价 * key_compare_cost (default 0.1) 键比较的代价,例如排序 * disk_temptable_create_cost (default 40.0) 内部myisam或innodb临时表的创建代价 * disk_temptable_row_cost (default 1.0) 内部myisam或innodb临时表的行代价 由上可以看出创建临时表的代价是很高的,尤其是内部的myisam或innodb临时表。 ## engine_cost * io_block_read_cost (default 1.0) 从磁盘读数据的代价,对innodb来说,表示从磁盘读一个page的代价 * memory_block_read_cost (default 1.0) 从内存读数据的代价,对innodb来说,表示从buffer pool读一个page的代价 目前io_block_read_cost和memory_block_read_cost默认值均为1,实际生产中建议酌情调大memory_block_read_cost,特别是对普通硬盘的场景。 ## 代价配置 cost参数可以通过修改mysql.server_cost和mysql.engine_cost来实现。初始这两个表中的记录cost_value项均为NULL, 代价值都取上两节介绍的初始值。 当修改cost_value为非NULL时,代价值按设定的值计算。修改方法如下: ~~~ ## 修改io_block_read_cost值为2 UPDATE mysql.engine_cost SET cost_value = 2.0 WHERE cost_name = 'io_block_read_cost'; # FLUSH OPTIMIZER_COSTS 生效,只对新连接有效,老连接无效。 FLUSH OPTIMIZER_COSTS; ~~~ 另外,在主备环境下,修改cost参数时主备都要修改。因为mysql.server_cost和mysql.engine_cost的更新不会参与复制。 ## 代价分析示例 初始化数据 ~~~ create table t1(c1 int primary key, c2 int unique,c3 int) engine=innodb; let $loop=100; while($loop) { eval insert into t1(c1,c2,c3) values($loop, $loop+1, $loop+2); dec $loop; } set optimizer_trace = "enabled=on"; ~~~ cost参数都取默认值,以下示例中会用到row_evaluate_cost(0.2),io_block_read_cost(1.0),io_block_read_cost(1.0),memory_block_read_cost(1.0) ### 示例1 以下语句选择覆盖索引c2 ~~~ explain select c1,c2 from t1 where c2 > 10; id select_type table partitions type possible_keys key key_len ref rows filtered Extra 1 SIMPLE t1 NULL range c2 c2 5 NULL 91 100.00 Using where; Using index ~~~ 查看optimizer_trace, 可以看出全表扫描代价为23.1,通过c2上的索引扫描代价为19.309, 最后选择c2上的索引扫描。 ~~~ "rows_estimation": [ { "table": "`t1`", "range_analysis": { "table_scan": { "rows": 100, "cost": 23.1 }, "potential_range_indexes": [ { "index": "PRIMARY", "usable": false, "cause": "not_applicable" }, { "index": "c2", "usable": true, "key_parts": [ "c2" ] } ], "best_covering_index_scan": { "index": "c2", "cost": 21.109, "chosen": true }, "setup_range_conditions": [ ], "group_index_range": { "chosen": false, "cause": "not_group_by_or_distinct" }, "analyzing_range_alternatives": { "range_scan_alternatives": [ { "index": "c2", "ranges": [ "10 < c2" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": true, "rows": 91, "cost": 19.309, "chosen": true } ], "analyzing_roworder_intersect": { "usable": false, "cause": "too_few_roworder_scans" } }, "chosen_range_access_summary": { "range_access_plan": { "type": "range_scan", "index": "c2", "rows": 91, "ranges": [ "10 < c2" ] }, "rows_for_plan": 91, "cost_for_plan": 19.309, "chosen": true } } } ] }, { "considered_execution_plans": [ { "plan_prefix": [ ], "table": "`t1`", "best_access_path": { "considered_access_paths": [ { "rows_to_scan": 91, "access_type": "range", "range_details": { "used_index": "c2" }, "resulting_rows": 91, "cost": 37.509, "chosen": true } ] }, "condition_filtering_pct": 100, "rows_for_plan": 91, "cost_for_plan": 37.509, "chosen": true } ] ~~~ 全表扫描的代价23.1 包括io和cpu的代价 ~~~ test_quick_select: double scan_time= cost_model->row_evaluate_cost(static_cast<double>(records)) + 1; Cost_estimate cost_est= head->file->table_scan_cost(); cost_est.add_io(1.1);//这里加1.1应该是个调节值 cost_est.add_cpu(scan_time); ~~~ 其中io代价table_scan_cost会根据buffer pool大小和索引大小来估算page in memory和in disk的比例,分别算出代价。 ~~~ handler::table_scan_cost() ha_innobase::scan_time()*table->cost_model()->page_read_cost(1.0);//1*1=1 //其中scan_time计算数据所占page数, ~~~ page_read_cost计算读取单个page的代价 ~~~ buffer_block_read_cost(pages_in_mem) + io_block_read_cost(pages_on_disk); ~~~ io代价为1+1.1=2.1 cpu代价为row_evaluate_cost ~~~ double row_evaluate_cost(double rows) const { DBUG_ASSERT(m_initialized); DBUG_ASSERT(rows >= 0.0); return rows * m_server_cost_constants->row_evaluate_cost(); // 100 * 0.2(row_evaluate_cost)=20; } ~~~ cpu代价为20+1=21; 最终代价为2.1+21=23.1 c2索引扫描代价19.309 同样也分为io和cpu代价 ~~~ multi_range_read_info_const: *cost= index_scan_cost(keyno, static_cast<double>(n_ranges), static_cast<double>(total_rows)); cost->add_cpu(cost_model->row_evaluate_cost(static_cast<double>(total_rows)) + 0.01); ~~~ io代价 1.0987925356750823*1=1.0987925356750823 ~~~ index_scan_cost: const double io_cost= index_only_read_time(index, rows) * //估算index占page个数 = 1.0987925356750823 table->cost_model()->page_read_cost_index(index, 1.0); //根据buffer pool大小和索引大小来估算page in memory和in disk的比例,计算读一个page的代价。 = 1 ~~~ cpu代价91*0.2+0.01=18.21 ~~~ cost->add_cpu(cost_model->row_evaluate_cost( static_cast<double>(total_rows)) + 0.01); //这里根据过滤条件算出的total_rows为91 ~~~ 最终代价1.0987925356750823+18.21=19.309 ### 示例2 以下语句选择了全表扫描 ~~~ explain select * from t1 where c2 > 10; id select_type table partitions type possible_keys key key_len ref rows filtered Extra 1 SIMPLE t1 NULL ALL c2 NULL NULL NULL 100 91.00 Using where ~~~ 查看optimizer_trace, 可以看出全表扫描代价为23.1,通过c2上的索引扫描代价为110.21, 最后选择全表扫描。 ~~~ "rows_estimation": [ { "table": "`t1`", "range_analysis": { "table_scan": { "rows": 100, "cost": 23.1 }, "potential_range_indexes": [ { "index": "PRIMARY", "usable": false, "cause": "not_applicable" }, { "index": "c2", "usable": true, "key_parts": [ "c2" ] } ], "setup_range_conditions": [ ], "group_index_range": { "chosen": false, "cause": "not_group_by_or_distinct" }, "analyzing_range_alternatives": { "range_scan_alternatives": [ { "index": "c2", "ranges": [ "10 < c2" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 91, "cost": 110.21, "chosen": false, "cause": "cost" } ], "analyzing_roworder_intersect": { "usable": false, "cause": "too_few_roworder_scans" } } } } ] }, { "considered_execution_plans": [ { "plan_prefix": [ ], "table": "`t1`", "best_access_path": { "considered_access_paths": [ { "rows_to_scan": 100, "access_type": "scan", "resulting_rows": 91, "cost": 21, "chosen": true } ] }, "condition_filtering_pct": 100, "rows_for_plan": 91, "cost_for_plan": 21, "chosen": true } ] }, ~~~ 全表扫描代价23.1 同上一节分析 c2索引扫描代价为110.21 上一节通过c2索引扫描代价为19.309,因为是覆盖索引不需要回表,所以代价较少。而此例是需要回表的。 ~~~ multi_range_read_info_const: *cost= read_cost(keyno, static_cast<double>(n_ranges), static_cast<double>(total_rows)); cost->add_cpu(cost_model->row_evaluate_cost( static_cast<double>(total_rows)) + 0.01); ~~~ io代价需回表 ~~~ read_cost: //92*1=92 const double io_cost= read_time(index, static_cast<uint>(ranges) static_cast<ha_rows>(rows)) * table->cost_model()->page_read_cost(1.0); read_time: //91+1=92 virtual double read_time(uint index, uint ranges, ha_rows rows) { return rows2double(ranges+rows); } ~~~ 这里回表时计算代价为每行代价为1,默认认为回表时每行都对于聚集索引的一个page. io代价为92 cpu代价为91*0.2+0.01=18.21 `cost->add_cpu(cost_model->row_evaluate_cost( static_cast<double>(total_rows)) + 0.01);` 最后代价为92+18.21=110.21 ## 总结 5.7 代价模型优化还在持续改进中,相信后续的版本会越来越好。代价的参数的配置需谨慎,需要大量的测试和验证。