Commit 50f29b0e authored by sergefp@mysql.com's avatar sergefp@mysql.com

Post-merge fixes and code cleanup

parent b1f07034
...@@ -7,7 +7,7 @@ filler char(200), ...@@ -7,7 +7,7 @@ filler char(200),
filler2 char(200), filler2 char(200),
index(key1), index(key1),
index(key2), index(key2),
) type=bdb; ) engine=bdb;
select * from t1 where (key1 >= 2 and key1 <= 10) or (pk >= 4 and pk <=8 ); select * from t1 where (key1 >= 2 and key1 <= 10) or (pk >= 4 and pk <=8 );
pk key1 key2 filler filler2 pk key1 key2 filler filler2
2 2 2 filler-data filler-data-2 2 2 2 filler-data filler-data-2
......
...@@ -7,7 +7,7 @@ filler char(200), ...@@ -7,7 +7,7 @@ filler char(200),
filler2 char(200), filler2 char(200),
index(key1), index(key1),
index(key2), index(key2),
) type=innodb; ) engine=innodb;
select * from t1 where (key1 >= 2 and key1 <= 10) or (pk >= 4 and pk <=8 ); select * from t1 where (key1 >= 2 and key1 <= 10) or (pk >= 4 and pk <=8 );
pk key1 key2 filler filler2 pk key1 key2 filler filler2
2 2 2 filler-data filler-data-2 2 2 2 filler-data filler-data-2
......
...@@ -15,7 +15,7 @@ create table t1 ( ...@@ -15,7 +15,7 @@ create table t1 (
filler2 char(200), filler2 char(200),
index(key1), index(key1),
index(key2), index(key2),
) type=bdb; ) engine=bdb;
--disable_query_log --disable_query_log
......
...@@ -15,7 +15,7 @@ create table t1 ( ...@@ -15,7 +15,7 @@ create table t1 (
filler2 char(200), filler2 char(200),
index(key1), index(key1),
index(key2), index(key2),
) type=innodb; ) engine=innodb;
--disable_query_log --disable_query_log
......
...@@ -274,7 +274,11 @@ class SEL_TREE :public Sql_alloc ...@@ -274,7 +274,11 @@ class SEL_TREE :public Sql_alloc
public: public:
enum Type { IMPOSSIBLE, ALWAYS, MAYBE, KEY, KEY_SMALLER } type; enum Type { IMPOSSIBLE, ALWAYS, MAYBE, KEY, KEY_SMALLER } type;
SEL_TREE(enum Type type_arg) :type(type_arg) {} SEL_TREE(enum Type type_arg) :type(type_arg) {}
SEL_TREE() :type(KEY) { keys_map.clear_all(); bzero((char*) keys,sizeof(keys));} SEL_TREE() :type(KEY)
{
keys_map.clear_all();
bzero((char*) keys,sizeof(keys));
}
SEL_ARG *keys[MAX_KEY]; SEL_ARG *keys[MAX_KEY];
key_map keys_map; /* bitmask of non-NULL elements in keys */ key_map keys_map; /* bitmask of non-NULL elements in keys */
List<SEL_IMERGE> merges; /* possible ways to read rows using index_merge */ List<SEL_IMERGE> merges; /* possible ways to read rows using index_merge */
...@@ -306,7 +310,8 @@ static ha_rows check_quick_keys(PARAM *param,uint index,SEL_ARG *key_tree, ...@@ -306,7 +310,8 @@ static ha_rows check_quick_keys(PARAM *param,uint index,SEL_ARG *key_tree,
char *max_key, uint max_key_flag); char *max_key, uint max_key_flag);
QUICK_RANGE_SELECT *get_quick_select(PARAM *param,uint index, QUICK_RANGE_SELECT *get_quick_select(PARAM *param,uint index,
SEL_ARG *key_tree, MEM_ROOT *alloc = NULL); SEL_ARG *key_tree,
MEM_ROOT *alloc = NULL);
static int get_quick_select_params(SEL_TREE *tree, PARAM *param, static int get_quick_select_params(SEL_TREE *tree, PARAM *param,
key_map& needed_reg, key_map& needed_reg,
bool index_read_can_be_used, bool index_read_can_be_used,
...@@ -338,7 +343,8 @@ bool get_quick_keys(PARAM *param,QUICK_RANGE_SELECT *quick,KEY_PART *key, ...@@ -338,7 +343,8 @@ bool get_quick_keys(PARAM *param,QUICK_RANGE_SELECT *quick,KEY_PART *key,
static bool eq_tree(SEL_ARG* a,SEL_ARG *b); static bool eq_tree(SEL_ARG* a,SEL_ARG *b);
static SEL_ARG null_element(SEL_ARG::IMPOSSIBLE); static SEL_ARG null_element(SEL_ARG::IMPOSSIBLE);
static bool null_part_in_key(KEY_PART *key_part, const char *key, uint length); static bool null_part_in_key(KEY_PART *key_part, const char *key,
uint length);
bool sel_trees_can_be_ored(SEL_TREE *tree1, SEL_TREE *tree2, PARAM* param); bool sel_trees_can_be_ored(SEL_TREE *tree1, SEL_TREE *tree2, PARAM* param);
...@@ -620,7 +626,7 @@ QUICK_SELECT_I::QUICK_SELECT_I() ...@@ -620,7 +626,7 @@ QUICK_SELECT_I::QUICK_SELECT_I()
QUICK_RANGE_SELECT::QUICK_RANGE_SELECT(THD *thd, TABLE *table, uint key_nr, QUICK_RANGE_SELECT::QUICK_RANGE_SELECT(THD *thd, TABLE *table, uint key_nr,
bool no_alloc, MEM_ROOT *parent_alloc) bool no_alloc, MEM_ROOT *parent_alloc)
:dont_free(0),error(0),range(0),cur_range(NULL) :dont_free(0),error(0),cur_range(NULL),range(0)
{ {
index= key_nr; index= key_nr;
head= table; head= table;
...@@ -654,9 +660,10 @@ QUICK_RANGE_SELECT::~QUICK_RANGE_SELECT() ...@@ -654,9 +660,10 @@ QUICK_RANGE_SELECT::~QUICK_RANGE_SELECT()
} }
QUICK_INDEX_MERGE_SELECT::QUICK_INDEX_MERGE_SELECT(THD *thd_param, TABLE *table) QUICK_INDEX_MERGE_SELECT::QUICK_INDEX_MERGE_SELECT(THD *thd_param,
:cur_quick_it(quick_selects), thd(thd_param), unique(NULL), TABLE *table)
pk_quick_select(NULL) :cur_quick_it(quick_selects),pk_quick_select(NULL),unique(NULL),
thd(thd_param)
{ {
index= MAX_KEY; index= MAX_KEY;
head= table; head= table;
...@@ -1231,7 +1238,7 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg, ...@@ -1231,7 +1238,7 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg,
double tree_read_time; double tree_read_time;
ha_rows tree_records; ha_rows tree_records;
bool pk_is_clustered= param->table->file->primary_key_is_clustered(); bool pk_is_clustered= param->table->file->primary_key_is_clustered();
bool have_cpk_scan; bool have_cpk_scan= false;
ha_rows records_for_unique= 0; ha_rows records_for_unique= 0;
ha_rows cpk_records= 0; ha_rows cpk_records= 0;
...@@ -1296,6 +1303,7 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg, ...@@ -1296,6 +1303,7 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg,
records_for_unique += tree_records; records_for_unique += tree_records;
} }
} }
DBUG_PRINT("info",("index_merge cost of index reads: %g", imerge_cost));
if (imerge_too_expensive) if (imerge_too_expensive)
DBUG_RETURN(1); DBUG_RETURN(1);
...@@ -1330,32 +1338,36 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg, ...@@ -1330,32 +1338,36 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg,
{ {
double n_blocks= double n_blocks=
ceil((double)(longlong)param->table->file->data_file_length / IO_SIZE); ceil((double)(longlong)param->table->file->data_file_length / IO_SIZE);
/* Don't ceil the following as it is an estimate */
double busy_blocks= double busy_blocks=
n_blocks * (1.0 - pow(1.0 - 1.0/n_blocks, records_for_unique)); n_blocks * (1.0 - pow(1.0 - 1.0/n_blocks, records_for_unique));
JOIN *join= param->thd->lex->select_lex.join; JOIN *join= param->thd->lex->select_lex.join;
if (!join || join->tables == 1) if (!join || join->tables == 1)
{
imerge_cost += busy_blocks*(DISK_SEEK_BASE_COST + imerge_cost += busy_blocks*(DISK_SEEK_BASE_COST +
DISK_SEEK_PROP_COST*n_blocks/busy_blocks); DISK_SEEK_PROP_COST*n_blocks/busy_blocks);
}
else else
{ {
/* /*
It can be a join with source table being non-last table, so assume It can be a join with source table being non-last table, so assume
that disk seeks are random here. that disk seeks are random here.
(TODO it is possible to determine if this *is* a last table in 'index (TODO it is possible to determine if this is a last table in 'index
checked for each record'-type join) checked for each record' join)
*/ */
imerge_cost += busy_blocks; imerge_cost += busy_blocks;
} }
} }
DBUG_PRINT("info",("index_merge cost with rowid-to-row scan: %g", imerge_cost));
/* PHASE 3: Add Unique operations cost */ /* PHASE 3: Add Unique operations cost */
imerge_cost += Unique::get_use_cost(param->mem_root, records_for_unique, double unique_cost=
Unique::get_use_cost(param->mem_root, records_for_unique,
param->table->file->ref_length, param->table->file->ref_length,
param->thd->variables.sortbuff_size); param->thd->variables.sortbuff_size);
if (unique_cost < 0.0)
DBUG_RETURN(1);
imerge_cost += unique_cost;
DBUG_PRINT("info",("index_merge total cost: %g", imerge_cost));
if (imerge_cost < *read_time) if (imerge_cost < *read_time)
{ {
*read_time= imerge_cost; *read_time= imerge_cost;
...@@ -1382,7 +1394,8 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg, ...@@ -1382,7 +1394,8 @@ static int get_index_merge_params(PARAM *param, key_map& needed_reg,
key blocks are half full (normally things are much better). key blocks are half full (normally things are much better).
*/ */
inline double get_index_only_read_time(PARAM* param, ha_rows records, int keynr) inline double get_index_only_read_time(PARAM* param, ha_rows records,
int keynr)
{ {
double read_time; double read_time;
uint keys_per_block= (param->table->file->block_size/2/ uint keys_per_block= (param->table->file->block_size/2/
...@@ -1445,12 +1458,7 @@ static int get_quick_select_params(SEL_TREE *tree, PARAM *param, ...@@ -1445,12 +1458,7 @@ static int get_quick_select_params(SEL_TREE *tree, PARAM *param,
read_index_only && read_index_only &&
(param->table->file->index_flags(keynr) & HA_KEY_READ_ONLY)) (param->table->file->index_flags(keynr) & HA_KEY_READ_ONLY))
{ {
/* /* We can resolve this by only reading through this key. */
We can resolve this by only reading through this key.
Assume that we will read trough the whole key range
and that all key blocks are half full (normally things are
much better).
*/
found_read_time=get_index_only_read_time(param, found_records, keynr); found_read_time=get_index_only_read_time(param, found_records, keynr);
} }
else else
...@@ -3893,7 +3901,8 @@ static void print_quick_sel_imerge(QUICK_INDEX_MERGE_SELECT *quick, ...@@ -3893,7 +3901,8 @@ static void print_quick_sel_imerge(QUICK_INDEX_MERGE_SELECT *quick,
DBUG_VOID_RETURN; DBUG_VOID_RETURN;
} }
void print_quick_sel_range(QUICK_RANGE_SELECT *quick,const key_map *needed_reg) void print_quick_sel_range(QUICK_RANGE_SELECT *quick,
const key_map *needed_reg)
{ {
QUICK_RANGE *range; QUICK_RANGE *range;
char buf[MAX_KEY/8+1]; char buf[MAX_KEY/8+1];
......
...@@ -76,11 +76,13 @@ Unique::Unique(qsort_cmp2 comp_func, void * comp_func_fixed_arg, ...@@ -76,11 +76,13 @@ Unique::Unique(qsort_cmp2 comp_func, void * comp_func_fixed_arg,
#define M_PI 3.14159265358979323846 #define M_PI 3.14159265358979323846
#endif #endif
#define M_E (exp(1)) #ifndef M_E
#define M_E (exp((double)1.0))
#endif
inline double log2_n_fact(double x) inline double log2_n_fact(double x)
{ {
return (2 * ( ((x)+1) * log(((x)+1)/M_E) + log(2*M_PI*((x)+1))/2 ) / log(2)); return (2 * (((x)+1)*log(((x)+1)/M_E) + log(2*M_PI*((x)+1))/2 ) / log(2));
} }
/* /*
...@@ -174,7 +176,8 @@ static double get_merge_many_buffs_cost(MEM_ROOT *alloc, ...@@ -174,7 +176,8 @@ static double get_merge_many_buffs_cost(MEM_ROOT *alloc,
key_size using max_in_memory_size memory. key_size using max_in_memory_size memory.
RETURN RETURN
Use cost as # of disk seeks. >=0 Cost in disk seeks.
<0 Out of memory.
NOTES NOTES
cost(using_unqiue) = cost(using_unqiue) =
...@@ -229,19 +232,28 @@ double Unique::get_use_cost(MEM_ROOT *alloc, uint nkeys, uint key_size, ...@@ -229,19 +232,28 @@ double Unique::get_use_cost(MEM_ROOT *alloc, uint nkeys, uint key_size,
result+= n_full_trees * log2_n_fact(max_elements_in_tree); result+= n_full_trees * log2_n_fact(max_elements_in_tree);
result /= TIME_FOR_COMPARE_ROWID; result /= TIME_FOR_COMPARE_ROWID;
/* Calculate cost of merging */ DBUG_PRINT("info",("unique trees sizes: %u=%u*%lu + %lu", nkeys,
n_full_trees, n_full_trees?max_elements_in_tree:0,
last_tree_elems));
if (!n_full_trees) if (!n_full_trees)
return result; return result;
/* There is more then one tree and merging is necessary. */ /*
/* Add cost of writing all trees to disk. */ There is more then one tree and merging is necessary.
First, add cost of writing all trees to disk.
*/
result += n_full_trees * ceil(key_size*max_elements_in_tree / IO_SIZE); result += n_full_trees * ceil(key_size*max_elements_in_tree / IO_SIZE);
result += ceil(key_size*last_tree_elems / IO_SIZE); result += ceil(key_size*last_tree_elems / IO_SIZE);
/* Cost of merge */ /* Cost of merge */
result += get_merge_many_buffs_cost(alloc, n_full_trees, double merge_cost= get_merge_many_buffs_cost(alloc, n_full_trees,
max_elements_in_tree, max_elements_in_tree,
last_tree_elems, key_size); last_tree_elems, key_size);
if (merge_cost < 0.0)
return merge_cost;
result += merge_cost;
/* /*
Add cost of reading the resulting sequence, assuming there were no Add cost of reading the resulting sequence, assuming there were no
duplicate elements. duplicate elements.
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment