- developers - lists.mariadb.org

[Maria-developers] Analysing MariaDB 5.5 sysbench performance regression
by Kristian Nielsen 31 Mar '12

31 Mar '12

I have been analysing the performance difference between MariaDB 5.5.21 and MySQL 5.5.21 using oltp.lua in sysbench 0.5. It turns out that MySQL has about 8% higher tps in this benchmark when running CPU bound. (See below for links to blog posts that started this). I took the base sysbench run and tweaked it to get stable numbers - basically put the datadir in ramdisk and adjusted my.cnf a bit (more details below). Then I used OProfile to analyse where the difference comes from. Oprofile is a really great tool for this kind of analysis, where CPU time spent is the measurement. It is very easy to use, and it gives detailed and accurate measurements on how much time is spent in each function. It is then easy to compare this against MySQL and find where we have performance regressions. Then further details is available by adding call graph to the reports, which shows from where the expensive parts are called; and by using the annotation facility which annotates every function and puts execution times on the individual statement and assembler functions. Here is an overview of where the extra time is spent. Time is in units of samples, which is 1/100000 of a clock cycle. Total was around 4.15M samples. Here are the top functions I found used more time in MariaDB, and the amount of extra samples spent in MariaDB: 39k mysql_select 20k alloc_root 16k JOIN::choose_subquery_plan 10k thd_ha_data 10k make_sortkey() 9k get_best_combination 9k handler::read_range_next 7k handler::ha_thd 7k update_ref_and_keys 7k Item::cmp_type 6k Item::operator delete 5k ha_maria::implicit_commit 4k end_write 4k sort_and_filter_keyuse 4k my_interval_timer 4k st_select_lex::master_unit 4k st_select_lex::save_leaf_tables 4k sel_arg_range_seq_next 4k thd_test_options 4k Item_func::walk 4k TABLE::init 4k sub_select 3k strnmov 3k ut_free 3k vio_write 3k ut_malloc_low 2k setup_tables 2k create_ref_for_key Some care is needed to interpret these data. GCC -O3 does heavy inlining. Oprofile is a statistical profiler, so numbers are somewhat uncertain, more so for smaller numbers. And modern CPUs are superscalar, out-of-order, and hyper-threaded, which makes interpreting individual instruction costs higly challenging. But the list can be used to drill down into the actual code; opannotate --source --assembler is very enlightening, in my experience (few examples below). >From this list, it seems that - not surprising - the 5.3 optimiser features seem to have added a bit of overhead. Much of this is spread out over many places in the code, though it could be that the root cause comes from few inlined functions. I checked mysql_select(), and most of the overhead is from lp:941889, expensive constructors in JOIN, which was inlined by GCC. I believe Sergey plans to fix this. More analysis of the individual functions is needed to find the cause there. It seems we are also using alloc_root() somewhat more. thd_ha_data and maybe handler_thd are likely due to using dynamically loaded innodb_plugin rather than built-in. I would like to do more analysis (and would probably need help from optimiser-knowledgable people for most of it) - I think it is feasible to remove most of the overhead in a modest time frame. But my main point is to present these results, to show that we have a very useful tool to pin-point these performance issues, and to start discussion in more detail (and I have plenty more of those). ----------------------------------------------------------------------- More background: These benchmarks started from these blog posts: http://www.mysqlperformanceblog.com/2012/02/18/mariadb-5-3-4-benchmarks/ http://blog.montyprogram.com/benchmarking-mariadb-5-3-4/ The main difference in these tests seem to be caused by different XtraDB buffer pool flush algorithm. So Axel tried using innodb_pluing in MariaDB instead, and got much smaller differences - but a difference was still there. So for 5.5 I took Axel's benchmark, adjusted until I got stable, fully cpu-bound numbers, and analysed the results, as above. I have the actual scripts and my.cnf available if anyone is interested. Main point is datadir in ramdisk, reducing redo log size to 300MB total, and running for 30 seconds. As examples, I analysed a handful of cases, and came up with two simple patches for specific issues. One is update_virtual_fields(), which originally accounted for 0.5% of total time in sysbench. This seems to be called for every row, so just the push/pop of registers in function prologue became significant. I fixed this with below patch; it puts a NULL pointer into the table instead of an empty list of virtual fields, and guards a few frequent calls with an if (NULL) check. Another is my_hash_sort_simple(). This is not a regression, but Oprofile shows we spend 10% of total time in this function. opannotate of the assembler showed the problem. Below simple patch speeds it up by 33%, by reducing 4 memory loads per loop to 2. Still, it would be nice to understand why this is called that much - here is the call stack: my_hash_sort_simple() ha_heap::write_row(). evaluate_join_record() sub_select() do_select() JOIN::exec() mysql_select() It would be nice to understand why this is needed - this is an expensive function, it needs to lookup every character in the collation tables. If this could be eliminated, that alone would probably be enough to win over MySQL, even without fixing any of the regressions. Some other top contenders in the profile are innodb functions. For example buf_calc_page_new_checksum(). This is expensive, as it needs to loop over every byte in each buffer pool page (16K), computing a 7-long dependency chain for each byte. This could be much faster if the checksum was done on 8-byte longwords, but AFAIK, it's not trivial to change the checksum algorithm due to the migration issues. Facebook may have done some work on this. Another top contender is row_sel_store_mysql_rec() and similar - they show high costs in the opannotate on loads from the records, this is probably cache misses from accessing pages in the buffer pool - not too much to do about that I think. Hope this helps, - Kristian. ----------------------------------------------------------------------- Patch for my_hash_sort_simple(): === modified file 'strings/ctype-simple.c' --- strings/ctype-simple.c 2012-01-13 14:50:02 +0000 +++ strings/ctype-simple.c 2012-03-16 16:34:31 +0000 @@ -306,19 +306,24 @@ void my_hash_sort_simple(CHARSET_INFO *c { register const uchar *sort_order=cs->sort_order; const uchar *end; + ulong n1, n2; /* Remove end space. We have to do this to be able to compare 'A ' and 'A' as identical */ end= skip_trailing_space(key, len); - + + n1= *nr1; + n2= *nr2; for (; key < (uchar*) end ; key++) { - nr1[0]^=(ulong) ((((uint) nr1[0] & 63)+nr2[0]) * - ((uint) sort_order[(uint) *key])) + (nr1[0] << 8); - nr2[0]+=3; + n1^=(ulong) ((((uint) n1 & 63)+n2) * + ((uint) sort_order[(uint) *key])) + (n1 << 8); + n2+=3; } + *nr1= n1; + *nr2= n2; } ----------------------------------------------------------------------- Patch for update_virtual_fields(): === modified file 'sql/filesort.cc' --- sql/filesort.cc 2012-03-11 22:45:18 +0000 +++ sql/filesort.cc 2012-03-16 19:17:58 +0000 @@ -589,7 +589,7 @@ static ha_rows find_all_keys(SORTPARAM * { if ((error= select->quick->get_next())) break; - if (!error) + if (!error && sort_form->vfield) update_virtual_fields(thd, sort_form); file->position(sort_form->record[0]); DBUG_EXECUTE_IF("debug_filesort", dbug_print_record(sort_form, TRUE);); @@ -608,7 +608,7 @@ static ha_rows find_all_keys(SORTPARAM * else { error=file->ha_rnd_next(sort_form->record[0]); - if (!error) + if (!error && sort_form->vfield) update_virtual_fields(thd, sort_form); if (!flag) { === modified file 'sql/records.cc' --- sql/records.cc 2012-01-13 14:50:02 +0000 +++ sql/records.cc 2012-03-16 19:15:51 +0000 @@ -344,7 +344,8 @@ static int rr_quick(READ_RECORD *info) break; } } - update_virtual_fields(info->thd, info->table); + if (info->table->vfield) + update_virtual_fields(info->thd, info->table); return tmp; } === modified file 'sql/sql_select.cc' --- sql/sql_select.cc 2012-03-11 22:45:18 +0000 +++ sql/sql_select.cc 2012-03-16 19:15:05 +0000 @@ -15824,7 +15824,8 @@ evaluate_join_record(JOIN *join, JOIN_TA DBUG_RETURN(NESTED_LOOP_KILLED); /* purecov: inspected */ } - update_virtual_fields(join->thd, join_tab->table); + if (join_tab->table->vfield) + update_virtual_fields(join->thd, join_tab->table); if (select_cond) { === modified file 'sql/table.cc' --- sql/table.cc 2012-03-02 23:03:20 +0000 +++ sql/table.cc 2012-03-18 16:10:18 +0000 @@ -2466,30 +2466,35 @@ int open_table_from_share(THD *thd, TABL /* Process virtual columns, if any. */ - if (!(vfield_ptr = (Field **) alloc_root(&outparam->mem_root, - (uint) ((share->vfields+1)* - sizeof(Field*))))) - goto err; - - outparam->vfield= vfield_ptr; - - for (field_ptr= outparam->field; *field_ptr; field_ptr++) + if (!share->vfields) + outparam->vfield= NULL; + else { - if ((*field_ptr)->vcol_info) + if (!(vfield_ptr = (Field **) alloc_root(&outparam->mem_root, + (uint) ((share->vfields+1)* + sizeof(Field*))))) + goto err; + + outparam->vfield= vfield_ptr; + + for (field_ptr= outparam->field; *field_ptr; field_ptr++) { - if (unpack_vcol_info_from_frm(thd, - outparam, - *field_ptr, - &(*field_ptr)->vcol_info->expr_str, - &error_reported)) + if ((*field_ptr)->vcol_info) { - error= 4; // in case no error is reported - goto err; + if (unpack_vcol_info_from_frm(thd, + outparam, + *field_ptr, + &(*field_ptr)->vcol_info->expr_str, + &error_reported)) + { + error= 4; // in case no error is reported + goto err; + } + *(vfield_ptr++)= *field_ptr; } - *(vfield_ptr++)= *field_ptr; } + *vfield_ptr= 0; // End marker } - *vfield_ptr= 0; // End marker #ifdef WITH_PARTITION_STORAGE_ENGINE if (share->partition_info_str_len && outparam->file) -----------------------------------------------------------------------

4 5

[Maria-developers] MariaDB 5.5.22-rc release
by Daniel Bartholomew 29 Mar '12

29 Mar '12

All, The MariaDB 5.5.22-rc release has been uploaded to the primary mirror and will be replicated to the other mirrors over the next few hours. The release notes and changelog pages for this release are at: https://kb.askmonty.org/en/mariadb-5522-release-notes https://kb.askmonty.org/en/mariadb-5522-changelog As soon as the release files are mirrored I will activate the downloads page, update various banners and links, and then I or Colin will send out an announce email and promote the release in the usual manner. In the mean time, please let me know if there is anything which should be mentioned in the release notes that currently is not. Thanks. -- Daniel Bartholomew MariaDB - http://mariadb.org Monty Program - http://montyprogram.com AskMonty Knowledgebase - http://kb.askmonty.org

1 0

[Maria-developers] WL#250 New (by Igor): Use the statistics from persistent statistical tables instead of the statistics provided by engine
by worklog-noreply＠askmonty.org 25 Mar '12

25 Mar '12

----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: Use the statistics from persistent statistical tables instead of the statistics provided by engine CREATION DATE..: Sun, 25 Mar 2012, 17:36 SUPERVISOR.....: IMPLEMENTOR....: Igor COPIES TO......: Igor Monty Psergey Sergei Timour CATEGORY.......: Server-Sprint TASK ID........: 250 (http://askmonty.org/worklog/?tid=250) VERSION........: Server-5.6 STATUS.........: Assigned PRIORITY.......: 60 WORKED HOURS...: 0 ESTIMATE.......: 0 (hours remain) ORIG. ESTIMATE.: 0 PROGRESS NOTES: DESCRIPTION: This task adds the functionality that allows optionally to use use the statistics from persistent statistical tables (see MWL #248) instead the statistics provided by engine. ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v4.0.0)

1 0

Re: [Maria-developers] [Commits] Improve filesort performance for small sorts
by Sergei Golubchik 24 Mar '12

24 Mar '12

Hi, Kristian! FYI: http://bazaar.launchpad.net/~mysql/mysql-server/trunk/revision/tor.didrikse… On Mar 24, knielsen(a)knielsen-hq.org wrote: > ------------------------------------------------------------ > revno: 3471 > revision-id: knielsen(a)knielsen-hq.org-20120324160859-6pzy6tcxvd6j4fbx > parent: monty(a)askmonty.org-20120323162239-h18wbqpbjz0p00jd > committer: knielsen(a)knielsen-hq.org > branch nick: mariadb-5.3 > timestamp: Sat 2012-03-24 17:08:59 +0100 > message: > Improve filesort performance for small sorts: Don't write pointers to records that we will never use. Regards, Sergei

1 0

[Maria-developers] Igor please review: [Commits] Rev 3464: BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view in file:///home/psergey/dev2/5.3-look60/]
by Sergei Petrunia 24 Mar '12

24 Mar '12

Hello Igor, Could you please review the below: ----- Forwarded message from Sergey Petrunya <psergey(a)askmonty.org> ----- From: Sergey Petrunya <psergey(a)askmonty.org> To: commits(a)mariadb.org X-Mailer: mail (GNU Mailutils 1.2) Date: Sat, 24 Mar 2012 01:40:02 +0400 (GST) Subject: [Commits] Rev 3464: BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view in file:///home/psergey/dev2/5.3-look60/ At file:///home/psergey/dev2/5.3-look60/ ------------------------------------------------------------ revno: 3464 revision-id: psergey(a)askmonty.org-20120323213956-ohxb9jitekne7yya parent: psergey(a)askmonty.org-20120321071820-t2azo8xxlm73mezd committer: Sergey Petrunya <psergey(a)askmonty.org> branch nick: 5.3-look60 timestamp: Sat 2012-03-24 01:39:56 +0400 message: BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view - Fix equality propagation to work with SJM nests and OR clauses (full descirption of problem and solution in the comment in the patch) === modified file 'mysql-test/r/subselect_sj2.result' --- a/mysql-test/r/subselect_sj2.result 2012-02-25 00:50:22 +0000 +++ b/mysql-test/r/subselect_sj2.result 2012-03-23 21:39:56 +0000 @@ -903,5 +903,48 @@ 3 1 1 4 1 1 DROP TABLE t1,t2; +# +# BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view +# +CREATE TABLE t1 ( +a VARCHAR(1), +b VARCHAR(1) NOT NULL, +KEY(a) +) ENGINE=InnoDB; +INSERT INTO t1 VALUES +('j','j'),('v','v'),('c','c'),('m','m'),('d','d'), +('y','y'),('t','t'),('d','d'),('s','s'),('r','r'), +('m','m'),('b','b'),('x','x'),('g','g'),('p','p'), +('q','q'),('w','w'),('d','d'),('e','e'); +CREATE ALGORITHM=TEMPTABLE VIEW v1 AS SELECT * FROM t1; +SELECT * FROM v1 +WHERE ( a, a ) IN ( +SELECT alias2.b, alias2.a +FROM t1 AS alias1, t1 AS alias2 +WHERE alias2.b = alias1.a +AND ( alias1.b >= alias1.a OR alias2.b = 'z' ) +); +a b +j j +v v +c c +m m +m m +d d +d d +d d +y y +t t +s s +r r +b b +x x +g g +p p +q q +w w +e e +DROP VIEW v1; +DROP TABLE t1; # This must be the last in the file: set optimizer_switch=@subselect_sj2_tmp; === modified file 'mysql-test/r/subselect_sj2_jcl6.result' --- a/mysql-test/r/subselect_sj2_jcl6.result 2012-02-25 00:50:22 +0000 +++ b/mysql-test/r/subselect_sj2_jcl6.result 2012-03-23 21:39:56 +0000 @@ -917,6 +917,49 @@ 3 1 1 4 1 1 DROP TABLE t1,t2; +# +# BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view +# +CREATE TABLE t1 ( +a VARCHAR(1), +b VARCHAR(1) NOT NULL, +KEY(a) +) ENGINE=InnoDB; +INSERT INTO t1 VALUES +('j','j'),('v','v'),('c','c'),('m','m'),('d','d'), +('y','y'),('t','t'),('d','d'),('s','s'),('r','r'), +('m','m'),('b','b'),('x','x'),('g','g'),('p','p'), +('q','q'),('w','w'),('d','d'),('e','e'); +CREATE ALGORITHM=TEMPTABLE VIEW v1 AS SELECT * FROM t1; +SELECT * FROM v1 +WHERE ( a, a ) IN ( +SELECT alias2.b, alias2.a +FROM t1 AS alias1, t1 AS alias2 +WHERE alias2.b = alias1.a +AND ( alias1.b >= alias1.a OR alias2.b = 'z' ) +); +a b +j j +v v +c c +m m +m m +d d +d d +d d +y y +t t +s s +r r +b b +x x +g g +p p +q q +w w +e e +DROP VIEW v1; +DROP TABLE t1; # This must be the last in the file: set optimizer_switch=@subselect_sj2_tmp; # === modified file 'mysql-test/r/subselect_sj2_mat.result' --- a/mysql-test/r/subselect_sj2_mat.result 2012-02-25 00:50:22 +0000 +++ b/mysql-test/r/subselect_sj2_mat.result 2012-03-23 21:39:56 +0000 @@ -905,6 +905,49 @@ 3 1 1 4 1 1 DROP TABLE t1,t2; +# +# BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view +# +CREATE TABLE t1 ( +a VARCHAR(1), +b VARCHAR(1) NOT NULL, +KEY(a) +) ENGINE=InnoDB; +INSERT INTO t1 VALUES +('j','j'),('v','v'),('c','c'),('m','m'),('d','d'), +('y','y'),('t','t'),('d','d'),('s','s'),('r','r'), +('m','m'),('b','b'),('x','x'),('g','g'),('p','p'), +('q','q'),('w','w'),('d','d'),('e','e'); +CREATE ALGORITHM=TEMPTABLE VIEW v1 AS SELECT * FROM t1; +SELECT * FROM v1 +WHERE ( a, a ) IN ( +SELECT alias2.b, alias2.a +FROM t1 AS alias1, t1 AS alias2 +WHERE alias2.b = alias1.a +AND ( alias1.b >= alias1.a OR alias2.b = 'z' ) +); +a b +j j +v v +c c +m m +m m +d d +d d +d d +y y +t t +s s +r r +b b +x x +g g +p p +q q +w w +e e +DROP VIEW v1; +DROP TABLE t1; # This must be the last in the file: set optimizer_switch=@subselect_sj2_tmp; set optimizer_switch=default; === modified file 'mysql-test/t/subselect_sj2.test' --- a/mysql-test/t/subselect_sj2.test 2012-01-19 19:44:43 +0000 +++ b/mysql-test/t/subselect_sj2.test 2012-03-23 21:39:56 +0000 @@ -1090,5 +1090,32 @@ DROP TABLE t1,t2; +--echo # +--echo # BUG#951937: Wrong result (missing rows) with semijoin+materialization, IN subquery, InnoDB, TEMPTABLE view +--echo # +CREATE TABLE t1 ( + a VARCHAR(1), + b VARCHAR(1) NOT NULL, + KEY(a) +) ENGINE=InnoDB; +INSERT INTO t1 VALUES +('j','j'),('v','v'),('c','c'),('m','m'),('d','d'), +('y','y'),('t','t'),('d','d'),('s','s'),('r','r'), +('m','m'),('b','b'),('x','x'),('g','g'),('p','p'), +('q','q'),('w','w'),('d','d'),('e','e'); + +CREATE ALGORITHM=TEMPTABLE VIEW v1 AS SELECT * FROM t1; + +# This query returns 6 rows instead of 19 + +SELECT * FROM v1 +WHERE ( a, a ) IN ( + SELECT alias2.b, alias2.a + FROM t1 AS alias1, t1 AS alias2 + WHERE alias2.b = alias1.a + AND ( alias1.b >= alias1.a OR alias2.b = 'z' ) +); +DROP VIEW v1; +DROP TABLE t1; --echo # This must be the last in the file: set optimizer_switch=@subselect_sj2_tmp; === modified file 'sql/item_cmpfunc.cc' --- a/sql/item_cmpfunc.cc 2012-02-28 14:41:55 +0000 +++ b/sql/item_cmpfunc.cc 2012-03-23 21:39:56 +0000 @@ -5388,7 +5388,7 @@ */ Item_equal::Item_equal(Item *f1, Item *f2, bool with_const_item) - : Item_bool_func(), eval_item(0), cond_false(0) + : Item_bool_func(), eval_item(0), cond_false(0), context_field(NULL) { const_item_cache= 0; with_const= with_const_item; @@ -5411,7 +5411,7 @@ */ Item_equal::Item_equal(Item_equal *item_equal) - : Item_bool_func(), eval_item(0), cond_false(0) + : Item_bool_func(), eval_item(0), cond_false(0), context_field(NULL) { const_item_cache= 0; List_iterator_fast<Item> li(item_equal->equal_items); === modified file 'sql/item_cmpfunc.h' --- a/sql/item_cmpfunc.h 2012-01-19 22:11:53 +0000 +++ b/sql/item_cmpfunc.h 2012-03-23 21:39:56 +0000 @@ -1700,9 +1700,16 @@ as datetimes. The comparator is used only if compare_as_dates=TRUE */ Arg_comparator cmp; + + /* + For Item_equal objects inside an OR clause: one of the fields that were + used in the original comparison. + */ + Item_field *context_field; public: inline Item_equal() - : Item_bool_func(), with_const(FALSE), eval_item(0), cond_false(0) + : Item_bool_func(), with_const(FALSE), eval_item(0), cond_false(0), + context_field(NULL) { const_item_cache=0 ;} Item_equal(Item *f1, Item *f2, bool with_const_item); Item_equal(Item_equal *item_equal); @@ -1729,6 +1736,8 @@ Item *transform(Item_transformer transformer, uchar *arg); virtual void print(String *str, enum_query_type query_type); CHARSET_INFO *compare_collation(); + + void set_context_field(Item_field *ctx_field) { context_field= ctx_field; } friend class Item_equal_fields_iterator; friend Item *eliminate_item_equal(COND *cond, COND_EQUAL *upper_levels, Item_equal *item_equal); === modified file 'sql/opt_subselect.cc' --- a/sql/opt_subselect.cc 2012-03-18 19:58:20 +0000 +++ b/sql/opt_subselect.cc 2012-03-23 21:39:56 +0000 @@ -171,6 +171,246 @@ */ +/* +EqualityPropagationAndSjmNests +****************************** + +The process of equality propagation has two parts +P1. Equality propagation +P2. Equality substitution [for a certain join order] + +The propagation part is not affected by SJM nests. In fact, it is done before +we determine the execution plan, i.e. before we even know we will use +SJM-nests for execution. + +The substitution part is affected. + +Substitution without SJMs +========================= +When one doesn't have SJM nests, tables have a strict join order: + + ---------------------------------> + t1 -- t2 -- t3 -- t4 --- t5 + + + ? ^ + \ + --(part-of-WHERE) + + +parts WHERE/ON and ref. expressions are attached at some point along the axis. +Expression is allowed to refer to a table column if the table is to the left of +the attachment point. For any given expression, we have a goal: + + "Move leftmost allowed attachment point as much as possible to the left" + +Substitution with SJMs - task setting +===================================== + +When SJM nests are present, there is no global strict table ordering anymore: + + + ---------------------------------> + + ot1 -- ot2 --- sjm -- ot4 --- ot5 + | + | Main execution + - - - - - - - - - - - - - - - - - - - - - - - - + | Materialization + it1 -- it2 --/ + + +Besides that, we must take into account that + - values for outer table columns, otN.col, are inaccessible at + materialization step (SJM-RULE) + - values for inner table columns, itN.col, are inaccessible at Main execution + step, except for SJ-Materialization-Scan and columns that are in the + subquery's select list. (SJM-RULE) + +Substitution with SJMs - solution +================================= + +First, we introduce global strict table ordering like this: + + ot1 - ot2 --\ /--- ot3 -- ot5 + \--- it1 --- it2 --/ + +Now, let's see how to meet (SJM-RULE). + +SJ-Materialization is only applicable for uncorrelated subqueries. From this, it +follows that any multiple equality will either +1. include only columns of outer tables, or +2. include only columns of inner tables, or +3. include columns of inner and outer tables, joined together through one + of IN-equalities. + +Cases #1 and #2 can be handled in the same way as with regular inner joins. + +Case #3 requires special handling, so that we don't construct violations of +(SJM-RULE). Let's consider possible ways to build violations. + +Equality propagation starts with the clause in this form + + top_query_where AND subquery_where AND in_equalities + +First, it builds multi-equalities. It can also build a mixed multi-equality + + multiple-equal(ot1.col, ot2.col, ... it1.col, itN.col) + +Multi-equalities are pushed down the OR-clauses in top_query_where and in +subquery_where, so it's possible that clauses like this one are built: + + subquery_cond OR (multiple-equal(it1.col, ot1.col,...) AND ...) + ^^^^^^^^^^^^^ \ + | this must be evaluated + \- can only be invaluated on the main phase. + on materialization phase + +Finally, equality substitution is started. It does two operations: + + +1. Field reference substitution +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +(In the code, this is Item_field::replace_equal_field) + +This is a process of replacing each reference to "tblX.col" +with the first element of the multi-equality. (REF-SUBST-ORIG) + +This behaviour can cause problems with Semi-join nests. Suppose, we have a +condition: + + func(it1.col, it2.col) + +and a multi-equality(ot1.col, it1.col). Then, reference to "it1.col" will be +replaced with "ot1.col", constructing a condition + + func(ot1.col, it2.col) + +which will be a vioation of (SJM-RULE). + +In order to avoid this, (REF-SUBST-ORIG) is amended as follows: + +- references to tables "itX.col" that are inner wrt some SJM nest, are + replaced with references to the first inner table from the same SJM nest. + +- references to top-level tables "otX.col" are replaced with references to + the first element of the multi-equality, no matter if that first element is + a column of a top-level table or of table from some SJM nest. + (REF-SUBST-SJM) + + The case where the first element is a table from an SJM nest $SJM is ok, + because it can be proven that $SJM uses SJ-Materialization-Scan, and + "unpacks" correct column values to the first element during the main + execution phase. + +2. Item_equal elimination +~~~~~~~~~~~~~~~~~~~~~~~~~ +(In the code: eliminate_item_equal) This is a process of taking + + multiple-equal(a,b,c,d,e) + +and replacing it with an equivalent expression which is an AND of pair-wise +equalities: + + a=b AND a=c AND ... + +The equalities are picked such that for any given join prefix (t1,t2...) the +subset of equalities that can be evaluated gives the most restrictive +filtering. + +Without SJM nests, it is sufficient to compare every multi-equality member +with the first one: + + elem1=elem2 AND elem1=elem3 AND elem1=elem4 ... + +When SJM nests are present, we should take care not to construct equalities +that violate the (SJM-RULE). This is achieved by generating separate sets of +equalites for top-level tables and for inner tables. That is, for the join +order + + ot1 - ot2 --\ /--- ot3 -- ot5 + \--- it1 --- it2 --/ + +we will generate + ot1.col=ot2.col + ot1.col=ot3.col + ot1.col=ot5.col + it2.col=it1.col + + +2.1 The problem with Item_equals and ORs +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +As has been mentioned above, multiple equalities are pushed down into OR +clauses, possibly building clauses like this: + + func(it.col2) OR multiple-equal(it1.col1, it1.col2, ot1.col) (1) + +where the first part of the clause has references to inner tables, while the +second has references to the top-level tables, which is a violation of +(SJM-RULE). + +AND-clauses of this kind do not create problems, because make_cond_for_table() +will take them apart. OR-clauses will not be split. It is possible to +split-out the part that's dependent on the inner table: + + func(it.col2) OR it1.col1=it1.col2 + +but this is a less-restrictive condition than condition (1). Current execution +scheme will still try to generate the "remainder" condition: + + func(it.col2) OR it1.col1=ot1.col + +which is a violation of (SJM-RULE). + +QQ: "ot1.col=it1.col" is checked at the upper level. Why was it not removed +here? +AA: because has a proper subset of conditions that are found on this level. + consider a join order of ot, sjm(it) + and a condition + ot.col=it.col AND ( ot.col=it.col='foo' OR it.col2='bar') + + we will produce: + table ot: nothing + table it: ot.col=it.col AND (ot.col='foo' OR it.col2='bar') + ^^^^ ^^^^^^^^^^^^^^^^ + | \ the problem is that + | this part condition didnt + | receive a substitution + | + +--- it was correct to subst, 'ot' is + the left-most. + + +Is there any merit in pushing inner=outer down into ORs? +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +Yes, there is. Consider the query: + + select * from ot + where ot.col in (select it.col from it where (it.col='foo' OR it.col='bar')) + +here, it may be useful to infer that + + (ot.col='foo' OR ot.col='bar') (CASE-FOR-SUBST) + +and attach that condition to the table 'ot'. + +Solution #1 +~~~~~~~~~~~ +Let make_cond_for_table() chop off the bad parts of produced OR clause. If it +has produced an OR clause: +- check if all of its branches are evaluable for the context join_tab +- If some aren't, destroy the OR clause + +Solution #2 +~~~~~~~~~~~ +Before the equality propagation phase, none of the OR clauses violate the +(SJM-RULE). This way, if we remember which tables the original equality +referred to, we can only generate equalities that refer to outer or inner +tables. Note that this will disallow handling of cases like (CASE-FOR-SUBST). +*/ + static bool subquery_types_allow_materialization(Item_in_subselect *in_subs); === modified file 'sql/sql_select.cc' --- a/sql/sql_select.cc 2012-03-13 20:49:18 +0000 +++ b/sql/sql_select.cc 2012-03-23 21:39:56 +0000 @@ -10666,6 +10666,9 @@ acceptable, as this happens rarely. The implementation without copying would be much more complicated. + For description of how equality propagation works with SJM nests, grep + for EqualityPropagationAndSjmNests. + @param left_item left term of the quality to be checked @param right_item right term of the equality to be checked @param item equality item if the equality originates from a condition @@ -10739,12 +10742,14 @@ { /* left_item_equal of an upper level contains left_item */ left_item_equal= new Item_equal(left_item_equal); + left_item_equal->set_context_field(((Item_field*) left_item)); cond_equal->current_level.push_back(left_item_equal); } if (right_copyfl) { /* right_item_equal of an upper level contains right_item */ right_item_equal= new Item_equal(right_item_equal); + right_item_equal->set_context_field(((Item_field*) right_item)); cond_equal->current_level.push_back(right_item_equal); } @@ -10830,6 +10835,7 @@ { item_equal= new Item_equal(item_equal); cond_equal->current_level.push_back(item_equal); + item_equal->set_context_field(field_item); } if (item_equal) { @@ -11480,6 +11486,8 @@ Item_equal::get_first() also takes similar measures for dealing with equality substitution in presense of SJM nests. + Grep for EqualityPropagationAndSjmNests for a more verbose description. + @return - The condition with generated simple equalities or a pointer to the simple generated equality, if success. @@ -11543,9 +11551,13 @@ on upper AND-levels. */ if (upper) - { + { + TABLE_LIST *native_sjm= embedding_sjm(item_equal->context_field); if (item_const && upper->get_const()) + { + /* Upper item also has "field_item=const". Don't produce equality here */ item= 0; + } else { Item_equal_fields_iterator li(*item_equal); @@ -11556,6 +11568,8 @@ break; } } + if (embedding_sjm(field_item) != native_sjm) + item= NULL; /* Don't produce equality */ } bool produce_equality= test(item == field_item); _______________________________________________ commits mailing list commits(a)mariadb.org https://lists.askmonty.org/cgi-bin/mailman/listinfo/commits ----- End forwarded message ----- -- BR Sergei -- Sergei Petrunia, Software Developer Monty Program AB, http://askmonty.org Blog: http://s.petrunia.net/blog

2 1

[Maria-developers] Autolinks to JIRA from the KB
by Daniel Bartholomew 21 Mar '12

21 Mar '12

All, I've created some autolinks for referring to JIRA tasks in the AskMonty Knowledgebase. I've documented how to use them on the "JIRA - project planning and tracking" page, but I wanted to make everyone here aware of them. http://kb.askmonty.org/en/jira-project-planning-and-tracking Basically, when you use mdev-xxx or todo-xxx (where xxx is the mdev or todo number) in a Knowledgebase article, the text will automatically be turned into a link to the task in JIRA. No need to copy the full url and paste it into the KB. Thanks. -- Daniel Bartholomew MariaDB - http://mariadb.org Monty Program - http://montyprogram.com AskMonty Knowledgebase - http://kb.askmonty.org

1 0

[Maria-developers] Need to lock temporary tables?
by webmonkeys＠alibi.com 20 Mar '12

20 Mar '12

>Description: InnoDB: Error: row_search_for_mysql() is called without ha_innobase::external_lock() >How-To-Repeat: create table foo ( data int ); create temporary table tmp_foo (select * from foo limit 0); insert into tmp_foo (data) values (1); lock tables foo write; # add ", tmp_foo read" to the above, to avoid the error. #"LOCK TABLES is permitted (but ignored) for a TEMPORARY table." # -- http://dev.mysql.com/doc/refman/5.0/en/lock-tables.html delete from foo; # The next statement causes # InnoDB: Error: row_search_for_mysql() is called without ha_innobase::external_lock() insert into foo (select * from tmp_foo); unlock tables; drop table tmp_foo; drop table foo; >Fix: Lock the temporary table, though mysql docs say you shouldn't have to. >Submitter-Id: <submitter ID> >Originator: >Organization: >MySQL support: [none | licence | email support | extended email support ] >Synopsis: temporary table wants a lock >Severity: non-critical >Priority: low >Category: mysql >Class: >Release: mysql-5.3.5-MariaDB-ga-mariadb113~squeeze ((MariaDB - http://mariadb.com/)) >C compiler: gcc-4.4.real (Debian 4.4.5-8) 4.4.5 >C++ compiler: g++-4.4.real (Debian 4.4.5-8) 4.4.5 >Environment: System: Linux machibashiri 2.6.32-5-amd64 #1 SMP Mon Mar 7 21:35:22 UTC 2011 x86_64 GNU/Linux Some paths: /usr/bin/perl /usr/bin/make /usr/bin/gcc /usr/bin/cc GCC: Using built-in specs. Target: x86_64-linux-gnu Configured with: ../src/configure -v --with-pkgversion='Debian 4.4.5-8' --with-bugurl=file:///usr/share/doc/gcc-4.4/README.Bugs --enable-languages=c,c++,fortran,objc,obj-c++ --prefix=/usr --program-suffix=-4.4 --enable-shared --enable-multiarch --enable-linker-build-id --with-system-zlib --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --with-gxx-include-dir=/usr/include/c++/4.4 --libdir=/usr/lib --enable-nls --enable-clocale=gnu --enable-libstdcxx-debug --enable-objc-gc --with-arch-32=i586 --with-tune=generic --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu Thread model: posix gcc version 4.4.5 (Debian 4.4.5-8) Compilation info (call): CC='gcc -static-libgcc' CFLAGS='-O2 -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized' CXX='g++ -static-libgcc' CXXFLAGS='-O2 -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized' LDFLAGS='' ASFLAGS='' Compilation info (used): CC='gcc -static-libgcc' CFLAGS=' -O2 -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized -DUNIV_LINUX -DUNIV_LINUX' CXX='g++ -static-libgcc' CXXFLAGS=' -O2 -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized -fno-implicit-templates -fno-exceptions -fno-rtti' LDFLAGS=' -rdynamic ' ASFLAGS='' LIBC: lrwxrwxrwx 1 root root 14 Apr 13 2011 /lib/libc.so.6 -> libc-2.11.2.so -rwxr-xr-x 1 root root 1432968 Jan 23 2011 /lib/libc-2.11.2.so -rw-r--r-- 1 root root 4433334 Jan 23 2011 /usr/lib/libc.a -rw-r--r-- 1 root root 247 Jan 23 2011 /usr/lib/libc.so Configure command: ../configure '--build=x86_64-linux-gnu' '--host=x86_64-linux-gnu' '--prefix=/usr' '--exec-prefix=/usr' '--libexecdir=/usr/sbin' '--datadir=/usr/share' '--localstatedir=/var/lib/mysql' '--includedir=/usr/include' '--infodir=/usr/share/info' '--mandir=/usr/share/man' '--with-server-suffix=-mariadb113~squeeze' '--with-comment=(MariaDB - http://mariadb.com/)' '--with-system-type=debian-linux-gnu' '--enable-shared' '--enable-static' '--enable-thread-safe-client' '--enable-assembler' '--enable-local-infile' '--with-pstack' '--with-fast-mutexes' '--with-unix-socket-path=/var/run/mysqld/mysqld.sock' '--with-mysqld-user=mysql' '--with-big-tables' '--without-docs' '--with-extra-charsets=all' '--with-libwrap' '--with-ssl=/usr' '--with-libevent' '--without-readline' '--with-zlib-dir=' '--with-embedded-server' '--with-embedded-privilege-control' '--with-plugins=max-no-ndb' 'build_alias=x86_64-linux-gnu' 'host_alias=x86_64-linux-gnu' 'CC=gcc -static-libgcc' 'CFLAGS=-O2! -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized' 'LDFLAGS=' 'CPPFLAGS=' 'CXX=g++ -static-libgcc' 'CXXFLAGS=-O2 -fno-omit-frame-pointer -g -pipe -Wall -Wno-uninitialized'

1 0

[Maria-developers] New KB Question: mariadb compilation flags
by Daniel Bartholomew 19 Mar '12

19 Mar '12

A new question has been posted to the Knowledgebase: http://kb.askmonty.org/en/mariadb-compilation-flags ---------------------------------------------------------------------- how come mariadb is compiled on linux 64 bit with -O2 rather than -O3? mysql is compiled with -O3. does it make any difference? thanks! ---------------------------------------------------------------------- Thanks. -- Daniel Bartholomew MariaDB - http://mariadb.org Monty Program - http://montyprogram.com AskMonty Knowledgebase - http://kb.askmonty.org

2 1

[Maria-developers] MariaDB 5.5.21
by Daniel Bartholomew 15 Mar '12

15 Mar '12

All, I've uploaded the MariaDB 5.5.21-beta packages to the primary mirror and within a few hours the files will be distributed to all of the mirrors. Over the past couple of weeks I've added several items to the release notes page for 5.5.21 and today I created the changelog page. The pages are here: http://kb.askmonty.org/en/mariadb-5521-release-notes http://kb.askmonty.org/en/mariadb-5521-changelog Once the files are distributed I will update the pages and add in the download links and remove the "unreleased" banners from the pages linked to from the release notes page. I will also send out an announcement email, and post announcements to twitter, FB, G+, and so on as usual. Please let me know ASAP if there is anything which is not on the release notes page which should be (or jump in and add it). Thanks. -- Daniel Bartholomew MariaDB - http://mariadb.org Monty Program - http://montyprogram.com AskMonty Knowledgebase - http://kb.askmonty.org

4 4

Re: [Maria-developers] mariadb 5.3.5 build on OpenBSD 5.0
by Sergei 15 Mar '12

15 Mar '12

Hi! On Mar 14, U Orsini wrote: > I know this isn't a good way of submitting reports, but I really don't > want to join mailing lists and teams just for this. Hope you can still > use this information. Some of it likely affects other *BSDs too. A good way of submitting bug reports is bugs.launchpad.net/maria You don't need to join a team to use it. I've submitted a bug report on your behalf: https://bugs.launchpad.net/maria/+bug/955760 Regards, Sergei

1 0