1
/* Copyright (C) 2001 MySQL AB
3
This program is free software; you can redistribute it and/or modify
4
it under the terms of the GNU General Public License as published by
5
the Free Software Foundation; version 2 of the License.
7
This program is distributed in the hope that it will be useful,
8
but WITHOUT ANY WARRANTY; without even the implied warranty of
9
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
10
GNU General Public License for more details.
12
You should have received a copy of the GNU General Public License
13
along with this program; if not, write to the Free Software
14
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA */
17
Function to handle quick removal of duplicates
18
This code is used when doing multi-table deletes to find the rows in
19
reference tables that needs to be deleted.
21
The basic idea is as follows:
23
Store first all strings in a binary tree, ignoring duplicates.
24
When the tree uses more memory than 'max_heap_table_size',
25
write the tree (in sorted order) out to disk and start with a new tree.
26
When all data has been generated, merge the trees (removing any found
29
The unique entries will be returned in sort order, to ensure that we do the
30
deletes in disk order.
33
#include "mysql_priv.h"
37
int unique_write_to_file(uchar* key, element_count count, Unique *unique)
40
Use unique->size (size of element stored in the tree) and not
41
unique->tree.size_of_element. The latter is different from unique->size
42
when tree implementation chooses to store pointer to key in TREE_ELEMENT
43
(instead of storing the element itself there)
45
return my_b_write(&unique->file, key, unique->size) ? 1 : 0;
48
int unique_write_to_ptrs(uchar* key, element_count count, Unique *unique)
50
memcpy(unique->record_pointers, key, unique->size);
51
unique->record_pointers+=unique->size;
55
Unique::Unique(qsort_cmp2 comp_func, void * comp_func_fixed_arg,
56
uint size_arg, ulonglong max_in_memory_size_arg)
57
:max_in_memory_size(max_in_memory_size_arg), size(size_arg), elements(0)
60
init_tree(&tree, (ulong) (max_in_memory_size / 16), 0, size, comp_func, 0,
61
NULL, comp_func_fixed_arg);
62
/* If the following fail's the next add will also fail */
63
my_init_dynamic_array(&file_ptrs, sizeof(BUFFPEK), 16, 16);
65
If you change the following, change it in get_max_elements function, too.
67
max_elements= (ulong) (max_in_memory_size /
68
ALIGN_SIZE(sizeof(TREE_ELEMENT)+size));
69
VOID(open_cached_file(&file, mysql_tmpdir,TEMP_PREFIX, DISK_BUFFER_SIZE,
78
Stirling's approximate formula is used:
80
n! ~= sqrt(2*M_PI*n) * (n/M_E)^n
82
Derivation of formula used for calculations is as follows:
84
log2(n!) = log(n!)/log(2) = log(sqrt(2*M_PI*n)*(n/M_E)^n) / log(2) =
86
= (log(2*M_PI*n)/2 + n*log(n/M_E)) / log(2).
89
inline double log2_n_fact(double x)
91
return (log(2*M_PI*x)/2 + x*log(x/M_E)) / M_LN2;
96
Calculate cost of merge_buffers function call for given sequence of
97
input stream lengths and store the number of rows in result stream in *last.
100
get_merge_buffers_cost()
101
buff_elems Array of #s of elements in buffers
102
elem_size Size of element stored in buffer
103
first Pointer to first merged element size
104
last Pointer to last merged element size
107
Cost of merge_buffers operation in disk seeks.
110
It is assumed that no rows are eliminated during merge.
111
The cost is calculated as
113
cost(read_and_write) + cost(merge_comparisons).
115
All bytes in the sequences is read and written back during merge so cost
116
of disk io is 2*elem_size*total_buf_elems/IO_SIZE (2 is for read + write)
118
For comparisons cost calculations we assume that all merged sequences have
119
the same length, so each of total_buf_size elements will be added to a sort
120
heap with (n_buffers-1) elements. This gives the comparison cost:
122
total_buf_elems* log2(n_buffers) / TIME_FOR_COMPARE_ROWID;
125
static double get_merge_buffers_cost(uint *buff_elems, uint elem_size,
126
uint *first, uint *last)
128
uint total_buf_elems= 0;
129
for (uint *pbuf= first; pbuf <= last; pbuf++)
130
total_buf_elems+= *pbuf;
131
*last= total_buf_elems;
133
int n_buffers= last - first + 1;
135
/* Using log2(n)=log(n)/log(2) formula */
136
return 2*((double)total_buf_elems*elem_size) / IO_SIZE +
137
total_buf_elems*log((double) n_buffers) / (TIME_FOR_COMPARE_ROWID * M_LN2);
142
Calculate cost of merging buffers into one in Unique::get, i.e. calculate
143
how long (in terms of disk seeks) the two calls
144
merge_many_buffs(...);
149
get_merge_many_buffs_cost()
150
buffer buffer space for temporary data, at least
151
Unique::get_cost_calc_buff_size bytes
152
maxbuffer # of full buffers
153
max_n_elems # of elements in first maxbuffer buffers
154
last_n_elems # of elements in last buffer
155
elem_size size of buffer element
158
maxbuffer+1 buffers are merged, where first maxbuffer buffers contain
159
max_n_elems elements each and last buffer contains last_n_elems elements.
161
The current implementation does a dumb simulation of merge_many_buffs
165
Cost of merge in disk seeks.
168
static double get_merge_many_buffs_cost(uint *buffer,
169
uint maxbuffer, uint max_n_elems,
170
uint last_n_elems, int elem_size)
173
double total_cost= 0.0;
174
uint *buff_elems= buffer; /* #s of elements in each of merged sequences */
177
Set initial state: first maxbuffer sequences contain max_n_elems elements
178
each, last sequence contains last_n_elems elements.
180
for (i = 0; i < (int)maxbuffer; i++)
181
buff_elems[i]= max_n_elems;
182
buff_elems[maxbuffer]= last_n_elems;
185
Do it exactly as merge_many_buff function does, calling
186
get_merge_buffers_cost to get cost of merge_buffers.
188
if (maxbuffer >= MERGEBUFF2)
190
while (maxbuffer >= MERGEBUFF2)
193
for (i = 0; i <= (int) maxbuffer - MERGEBUFF*3/2; i += MERGEBUFF)
195
total_cost+=get_merge_buffers_cost(buff_elems, elem_size,
197
buff_elems + i + MERGEBUFF-1);
200
total_cost+=get_merge_buffers_cost(buff_elems, elem_size,
202
buff_elems + maxbuffer);
207
/* Simulate final merge_buff call. */
208
total_cost += get_merge_buffers_cost(buff_elems, elem_size,
209
buff_elems, buff_elems + maxbuffer);
215
Calculate cost of using Unique for processing nkeys elements of size
216
key_size using max_in_memory_size memory.
219
Unique::get_use_cost()
220
buffer space for temporary data, use Unique::get_cost_calc_buff_size
221
to get # bytes needed.
222
nkeys #of elements in Unique
223
key_size size of each elements in bytes
224
max_in_memory_size amount of memory Unique will be allowed to use
231
cost(create_trees) + (see #1)
232
cost(merge) + (see #2)
233
cost(read_result) (see #3)
235
1. Cost of trees creation
236
For each Unique::put operation there will be 2*log2(n+1) elements
237
comparisons, where n runs from 1 tree_size (we assume that all added
238
elements are different). Together this gives:
240
n_compares = 2*(log2(2) + log2(3) + ... + log2(N+1)) = 2*log2((N+1)!)
242
then cost(tree_creation) = n_compares*ROWID_COMPARE_COST;
244
Total cost of creating trees:
245
(n_trees - 1)*max_size_tree_cost + non_max_size_tree_cost.
247
Approximate value of log2(N!) is calculated by log2_n_fact function.
250
If only one tree is created by Unique no merging will be necessary.
251
Otherwise, we model execution of merge_many_buff function and count
252
#of merges. (The reason behind this is that number of buffers is small,
253
while size of buffers is big and we don't want to loose precision with
256
3. If only one tree is created by Unique no disk io will happen.
257
Otherwise, ceil(key_len*n_keys) disk seeks are necessary. We assume
258
these will be random seeks.
261
double Unique::get_use_cost(uint *buffer, uint nkeys, uint key_size,
262
ulonglong max_in_memory_size)
264
ulong max_elements_in_tree;
265
ulong last_tree_elems;
266
int n_full_trees; /* number of trees in unique - 1 */
269
max_elements_in_tree= ((ulong) max_in_memory_size /
270
ALIGN_SIZE(sizeof(TREE_ELEMENT)+key_size));
272
n_full_trees= nkeys / max_elements_in_tree;
273
last_tree_elems= nkeys % max_elements_in_tree;
275
/* Calculate cost of creating trees */
276
result= 2*log2_n_fact(last_tree_elems + 1.0);
278
result+= n_full_trees * log2_n_fact(max_elements_in_tree + 1.0);
279
result /= TIME_FOR_COMPARE_ROWID;
281
DBUG_PRINT("info",("unique trees sizes: %u=%u*%lu + %lu", nkeys,
282
n_full_trees, n_full_trees?max_elements_in_tree:0,
289
There is more then one tree and merging is necessary.
290
First, add cost of writing all trees to disk, assuming that all disk
291
writes are sequential.
293
result += DISK_SEEK_BASE_COST * n_full_trees *
294
ceil(((double) key_size)*max_elements_in_tree / IO_SIZE);
295
result += DISK_SEEK_BASE_COST * ceil(((double) key_size)*last_tree_elems / IO_SIZE);
298
double merge_cost= get_merge_many_buffs_cost(buffer, n_full_trees,
299
max_elements_in_tree,
300
last_tree_elems, key_size);
301
if (merge_cost < 0.0)
304
result += merge_cost;
306
Add cost of reading the resulting sequence, assuming there were no
309
result += ceil((double)key_size*nkeys/IO_SIZE);
316
close_cached_file(&file);
318
delete_dynamic(&file_ptrs);
322
/* Write tree to disk; clear tree */
326
elements+= tree.elements_in_tree;
327
file_ptr.count=tree.elements_in_tree;
328
file_ptr.file_pos=my_b_tell(&file);
330
if (tree_walk(&tree, (tree_walk_action) unique_write_to_file,
331
(void*) this, left_root_right) ||
332
insert_dynamic(&file_ptrs, (uchar*) &file_ptr))
340
Clear the tree and the file.
341
You must call reset() if you want to reuse Unique after walk().
349
If elements != 0, some trees were stored in the file (see how
350
flush() works). Note, that we can not count on my_b_tell(&file) == 0
351
here, because it can return 0 right after walk(), and walk() does not
352
reset any Unique member.
356
reset_dynamic(&file_ptrs);
357
reinit_io_cache(&file, WRITE_CACHE, 0L, 0, 1);
363
The comparison function, passed to queue_init() in merge_walk() and in
364
merge_buffers() when the latter is called from Uniques::get() must
365
use comparison function of Uniques::tree, but compare members of struct
371
static int buffpek_compare(void *arg, uchar *key_ptr1, uchar *key_ptr2)
373
BUFFPEK_COMPARE_CONTEXT *ctx= (BUFFPEK_COMPARE_CONTEXT *) arg;
374
return ctx->key_compare(ctx->key_compare_arg,
375
*((uchar **) key_ptr1), *((uchar **)key_ptr2));
384
Function is very similar to merge_buffers, but instead of writing sorted
385
unique keys to the output file, it invokes walk_action for each key.
386
This saves I/O if you need to pass through all unique keys only once.
390
All params are 'IN' (but see comment for begin, end):
391
merge_buffer buffer to perform cached piece-by-piece loading
392
of trees; initially the buffer is empty
393
merge_buffer_size size of merge_buffer. Must be aligned with
395
key_length size of tree element; key_length * (end - begin)
396
must be less or equal than merge_buffer_size.
397
begin pointer to BUFFPEK struct for the first tree.
398
end pointer to BUFFPEK struct for the last tree;
399
end > begin and [begin, end) form a consecutive
400
range. BUFFPEKs structs in that range are used and
401
overwritten in merge_walk().
402
walk_action element visitor. Action is called for each unique
404
walk_action_arg argument to walk action. Passed to it on each call.
405
compare elements comparison function
406
compare_arg comparison function argument
407
file file with all trees dumped. Trees in the file
408
must contain sorted unique values. Cache must be
409
initialized in read mode.
415
static bool merge_walk(uchar *merge_buffer, ulong merge_buffer_size,
416
uint key_length, BUFFPEK *begin, BUFFPEK *end,
417
tree_walk_action walk_action, void *walk_action_arg,
418
qsort_cmp2 compare, void *compare_arg,
421
BUFFPEK_COMPARE_CONTEXT compare_context = { compare, compare_arg };
424
merge_buffer_size < (ulong) (key_length * (end - begin + 1)) ||
425
init_queue(&queue, (uint) (end - begin), offsetof(BUFFPEK, key), 0,
426
buffpek_compare, &compare_context))
428
/* we need space for one key when a piece of merge buffer is re-read */
429
merge_buffer_size-= key_length;
430
uchar *save_key_buff= merge_buffer + merge_buffer_size;
431
uint max_key_count_per_piece= (uint) (merge_buffer_size/(end-begin) /
433
/* if piece_size is aligned reuse_freed_buffer will always hit */
434
uint piece_size= max_key_count_per_piece * key_length;
435
uint bytes_read; /* to hold return value of read_to_buffer */
439
Invariant: queue must contain top element from each tree, until a tree
440
is not completely walked through.
441
Here we're forcing the invariant, inserting one element from each tree
444
for (top= begin; top != end; ++top)
446
top->base= merge_buffer + (top - begin) * piece_size;
447
top->max_keys= max_key_count_per_piece;
448
bytes_read= read_to_buffer(file, top, key_length);
449
if (bytes_read == (uint) (-1))
451
DBUG_ASSERT(bytes_read);
452
queue_insert(&queue, (uchar *) top);
454
top= (BUFFPEK *) queue_top(&queue);
455
while (queue.elements > 1)
458
Every iteration one element is removed from the queue, and one is
459
inserted by the rules of the invariant. If two adjacent elements on
460
the top of the queue are not equal, biggest one is unique, because all
461
elements in each tree are unique. Action is applied only to unique
464
void *old_key= top->key;
466
read next key from the cache or from the file and push it to the
467
queue; this gives new top.
469
top->key+= key_length;
470
if (--top->mem_count)
471
queue_replaced(&queue);
472
else /* next piece should be read */
474
/* save old_key not to overwrite it in read_to_buffer */
475
memcpy(save_key_buff, old_key, key_length);
476
old_key= save_key_buff;
477
bytes_read= read_to_buffer(file, top, key_length);
478
if (bytes_read == (uint) (-1))
480
else if (bytes_read > 0) /* top->key, top->mem_count are reset */
481
queue_replaced(&queue); /* in read_to_buffer */
485
Tree for old 'top' element is empty: remove it from the queue and
486
give all its memory to the nearest tree.
488
queue_remove(&queue, 0);
489
reuse_freed_buff(&queue, top, key_length);
492
top= (BUFFPEK *) queue_top(&queue);
493
/* new top has been obtained; if old top is unique, apply the action */
494
if (compare(compare_arg, old_key, top->key))
496
if (walk_action(old_key, 1, walk_action_arg))
501
Applying walk_action to the tail of the last tree: this is safe because
502
either we had only one tree in the beginning, either we work with the
503
last tree in the queue.
509
if (walk_action(top->key, 1, walk_action_arg))
511
top->key+= key_length;
513
while (--top->mem_count);
514
bytes_read= read_to_buffer(file, top, key_length);
515
if (bytes_read == (uint) (-1))
521
delete_queue(&queue);
528
Walks consecutively through all unique elements:
529
if all elements are in memory, then it simply invokes 'tree_walk', else
530
all flushed trees are loaded to memory piece-by-piece, pieces are
531
sorted, and action is called for each unique value.
532
Note: so as merging resets file_ptrs state, this method can change
533
internal Unique state to undefined: if you want to reuse Unique after
534
walk() you must call reset() first!
538
action function-visitor, typed in include/my_tree.h
539
function is called for each unique element
540
arg argument for visitor, which is passed to it on each call
546
bool Unique::walk(tree_walk_action action, void *walk_action_arg)
551
if (elements == 0) /* the whole tree is in memory */
552
return tree_walk(&tree, action, walk_action_arg, left_root_right);
554
/* flush current tree to the file to have some memory for merge buffer */
557
if (flush_io_cache(&file) || reinit_io_cache(&file, READ_CACHE, 0L, 0, 0))
559
if (!(merge_buffer= (uchar *) my_malloc((ulong) max_in_memory_size, MYF(0))))
561
res= merge_walk(merge_buffer, (ulong) max_in_memory_size, size,
562
(BUFFPEK *) file_ptrs.buffer,
563
(BUFFPEK *) file_ptrs.buffer + file_ptrs.elements,
564
action, walk_action_arg,
565
tree.compare, tree.custom_arg, &file);
566
my_free((char*) merge_buffer, MYF(0));
571
Modify the TABLE element so that when one calls init_records()
572
the rows will be read in priority order.
575
bool Unique::get(TABLE *table)
577
SORTPARAM sort_param;
578
table->sort.found_records=elements+tree.elements_in_tree;
580
if (my_b_tell(&file) == 0)
582
/* Whole tree is in memory; Don't use disk if you don't need to */
583
if ((record_pointers=table->sort.record_pointers= (uchar*)
584
my_malloc(size * tree.elements_in_tree, MYF(0))))
586
(void) tree_walk(&tree, (tree_walk_action) unique_write_to_ptrs,
587
this, left_root_right);
591
/* Not enough memory; Save the result to file && free memory used by tree */
595
IO_CACHE *outfile=table->sort.io_cache;
596
BUFFPEK *file_ptr= (BUFFPEK*) file_ptrs.buffer;
597
uint maxbuffer= file_ptrs.elements - 1;
602
/* Open cached file if it isn't open */
603
outfile=table->sort.io_cache=(IO_CACHE*) my_malloc(sizeof(IO_CACHE),
606
if (!outfile || (! my_b_inited(outfile) && open_cached_file(outfile,mysql_tmpdir,TEMP_PREFIX,READ_RECORD_BUFFER, MYF(MY_WME))))
608
reinit_io_cache(outfile,WRITE_CACHE,0L,0,0);
610
bzero((char*) &sort_param,sizeof(sort_param));
611
sort_param.max_rows= elements;
612
sort_param.sort_form=table;
613
sort_param.rec_length= sort_param.sort_length= sort_param.ref_length=
615
sort_param.keys= (uint) (max_in_memory_size / sort_param.sort_length);
616
sort_param.not_killable=1;
618
if (!(sort_buffer=(uchar*) my_malloc((sort_param.keys+1) *
619
sort_param.sort_length,
622
sort_param.unique_buff= sort_buffer+(sort_param.keys*
623
sort_param.sort_length);
625
sort_param.compare= (qsort2_cmp) buffpek_compare;
626
sort_param.cmp_context.key_compare= tree.compare;
627
sort_param.cmp_context.key_compare_arg= tree.custom_arg;
629
/* Merge the buffers to one file, removing duplicates */
630
if (merge_many_buff(&sort_param,sort_buffer,file_ptr,&maxbuffer,&file))
632
if (flush_io_cache(&file) ||
633
reinit_io_cache(&file,READ_CACHE,0L,0,0))
635
if (merge_buffers(&sort_param, &file, outfile, sort_buffer, file_ptr,
636
file_ptr, file_ptr+maxbuffer,0))
641
if (flush_io_cache(outfile))
644
/* Setup io_cache for reading */
645
save_pos=outfile->pos_in_file;
646
if (reinit_io_cache(outfile,READ_CACHE,0L,0,0))
648
outfile->end_of_file=save_pos;
added
removed
sql/uniques.cc
