Best Practices for Deploying MySQL on the Solaris Platform

Best Practices For Deploying MySQL on Solaris Presented by, MySQL AB® & O’Reilly Media, Inc. Ritu Kamboj Jenny Chen Agenda  MySQL -Solaris Integration  MySQL High Availability Data Service  Consolidate MySQL Deployment Using Solaris Containers  DTrace (Jenny Chen)‫‏‬  MySQL-Solaris Integration Optimized MySQL on Open Solaris  MySQL 5.0.45 (32bit) integrated with Open Solaris build 76  SXDE 01/08  MySQL 5.0.45 (64bit) integrated with Open Solaris build 87  SXCE Layout of MySQL on Open Solaris  MySQL 5.0.45 packages  Default data directory  Default configuration directory  /etc/mysql/5.0  Installation directory  /usr/mysql/5.0/  /var/mysql/5.0/data  Latest version accessible from /usr/mysql/bin  Symbolic link from all directories involved in 5.0.45 release to corresponding /usr/mysql Optimization Of MySQL On Open Solaris  Compiled with optimal compiler options using Sun Studio compilers  Added SMF support for MySQL  Easily accommodates varying configurations  Initializes MySQL database thereby enhancing user experience  Mysql user can manage MySQL database via SMF  Support for MySQL cluster engine (ndb)‫‏‬ Compiler Optimization  Enable in-lining  Change header file univ.i to add Sun Studio  If (!defined(SUNPRO_C)‫‏‬  #define UNIV_MUST_NOT_INLINE  Enable pre-fetching  -xprefetch=auto and -xprefetch_level=3  Set optimization level  -xO4 Compiler Optimization (Contd)‫‏‬  Feedback optimization (not yet implemented)‫‏‬  Workload is key  -xprofile -xipo  Interprocess optimization  About 10% improvement  Link with libmtmalloc  Library for threaded application  About 8% improvement Service Management Facility (SMF)‫‏‬  Makes Solaris Services Self-Healing  Services automatically restart in dependency order  Misbehaving and mis-configured services are easier to debug  Log files for each service  Administrators can securely delegate tasks to non-root users MySQL-Service Management Facility  Dynamically configured properties  mysql/bin  Installation , default : /usr/mysql/5.0/bin  mysql/data  Data Directory , Default: /var/mysql/5.0/data  mysql/enable_64bit  Flag to select 32bit or 64bit , Default : false  First time installations  Creates system tables (mysql_install_db.sh )‫‏‬ Starting MySQL on Open Solaris  Older Open Solaris Build (Build 76- Build-87)‫‏‬  32Bit MySQL integrated  Create mysql user  Enable MySQL SMF service  Latest Open Solaris Build (Build 87 onwards)‫‏‬  32Bit and 64bit MySQL integrated  Default mode : 32bit service  For starting 64bit  Set enable_64bit == true Optimized MySQL on Solaris 10  CoolStack  SAMP stack + more  Optimized MySQL download  Standalone package  SAMP stack component  Available at Sun Download center:  Version MySQL 5.0.45  Similar optimizations as in Open Solaris  http://cooltools.sunsource.net/coolstack  MySQL-High Availability Data Service  Provides general purpose HA platform Availability is our customers' most critical requirement Sun Cluster VOC Survey Solaris Cluster Overview Fifty percent of enterprises that lack a recovery plan go out of business within one year of a significant disaster Gartner Group Solaris Cluster : Hardware Components  Servers with local storage  Can have up to 16 nodes  Shared storage  Tolerates single-node failures  Centralizes configuration files  Cluster interconnect At least two redundant networks  Public network interfaces  Spreads outbound packets  Cluster membership monitor Solaris Cluster Algorithms  Ensures data integrity  Determines cluster membership  Cluster configuration repository  Global repository  Ensures consistent view  Disk Fencing  Fences off non-cluster nodes  Prevents Partition  Quorum  Uses a majority voting schema  HA-MySQL is a failover data service MySQL High Availability Data Service Node 1 Node 2 Node 3 Node 4 D B  Supported configurations  Standalone MySQL server  MySQL replication server MySQL High Availability Data Service  Single/Multiple MySQL instances in master configuration  Single/Multiple MySQL instances in slave configuration  Solaris containers support Global zone  Non-global failover zone  Non-global zone HA-MySQL Service Deployment MySQL Stor Node1 Host Node2 MySQL Stor Node1:Zone1 Host Node2:Zone2  Enhanced end-to-end infrastructure availability  Continuous MySQL Availability  Automatic failover if master node fails  Low cost solution  Software is free and open sourced  Efficient Resource Utilization  Multiple applications can be consolidated  Ease of operations  SC enables clustered systems to be managed as if they were on a single system MySQL-Solaris Cluster Benefits Additional information  Step by step deployment guide  http://docs.sun.com/app/docs/doc/819-3059  Failover study of HA-MySQL  http://blogs.sun.com/krishs/date/200804  Solaris Cluster  http://www.sun.com/software/solaris/cluster/index.xml http://opensolaris.org/os/community/ha-clusters/ Consolidate MySQL installations using Solaris Containers  Containers : Zones + Resource Management  Zones: isolated virtual application environments  Resource management – resource control (CPU, Memory)‫‏‬ Solaris Containers  Achieving Consolidation Goals  Reduce Hardware  Combine low utilization systems  Isolate applications from faults  Maintain Service Levels  Fine tune response times  System Configuration MySQL Consolidation Study Sun Fire X4100 (4 CPU, 8 GB Memory)‫‏‬  Local container configuration ( 1 CPU , 1 GB Memory  Sysbench Read-only Sysbench read-only 1900 1800 675 650 625 600 Sysbench read-write Sysbench read-write Throughput Throughput 1700 1600 1500 1400 1300 32 64 128 256 No Containers With Containers 575 550 525 500 475 450 425 32 64 128 256 No Containers With Containers Number of threads Number of threads  Dtrace (Jenny Chen)‫‏‬ DTrace  Use DTrace with MySQL to drill down Monitoring MySQL Performance  Examples & Solutions  Easy Steps To add DTrace probes into MySQL core server and storage Engines  Easy Steps to display MySQL DTrace probes into Chime visualization Tool for DTrace Why DTrace  Solaris 10 Dynamic Tracing Facility to provide comprehensive view of operating system and application behaviour > DTrace to examine particular system areas: disk I/O, CPU, Memory > Process Tracing and Debugging USDT(User-level statically defined tracing) place custom probes in application code Add USDT into MySQL source to monitor MySQL and gather the useful data missing by the current MySQL monitor tools: DTrace: Monitor I/Os  Exam I/O wait time by filename and mysqld(Available at DTraceToolkit)‫‏‬ #!/usr/sbin/dtrace –s #pragma D option quiet io:::wait-start /‫‏‬execname‫“‏==‏‬mysqld”‫/‏‬ { self->start = timestamp; } io:::wait-done /‫‏‬execname‫“‏==‏‬mysqld”‫‏&&‏‬self->start / { this->elapsed = timestamp - self->start; @files[pid,args[1]->dev_pathname, args[2]->fi_pathname] = sum(this->elapsed); self->start = 0; } profile:::tick-5s {‫‏‏‬printf(“-------------------------------------------------\n”); printf(“%6s %8s %20s %50s\n”,‫“‏‬PID”,‫“‏‬TIME”,‫“‏‬DEVICE”,‫“‏‬FILE”); printa(“%6d %@8d %20s %8s \n”,‫@‏‬files); printf‫\------------------------------------------------“(‏‬n”);‫‏‏‏‏‏‏}‏‬ #./mysqliowait.d --------------------------------------------------PID 113 TIME DEVICE 234 /dev/dsk/c0t2d0s0 FILE /usr/local/mysql/data/ibdata1 DTrace: Monitor CPU  classic performance problem #!/usr/sbin/dtrace -s #pragma D option quiet syscall:::entry { self->ts=vtimestamp; } syscall:::return /self->ts/ { @a[execname, probefunc] = count(); @b[execname, probefunc] = sum (vtimestamp - self->ts); self->ts=0; } END { printf("%-16s %-16s %-8s\n","EXEC","SYSCALL","COUNT"); printa("%-16s %-16s %-@8d\n",@a); printf("%-16s %-16s %-8s\n","EXEC","SYSCALL","TIME"); printa("%-16s %-16s %-@8d\n",@b); } # ./syscall.d EXEC SYSCALL COUNT devfsadm lwp_park 1 dtrace fstat 1 ... mysqld read 106542 mysqld gtime 109613 mysqld pread 1181669 --------------------------------------------------EXEC SYSCALL TIME dtrace lwp_sigmask 218 pkill getpid 302 ... mysqld read 259284183 mysqld write 267556239 mysqld pread 4650457224  Agrregated user stack backtrace to understand of the nature of pread() in MySQL source code – useful for mysql developers # dtrace -n 'syscall::pread:entry / execname == "mysqld" / { @[ustack()]=count() } dtrace: description 'syscall::pread:entry ' matched 1 probe ... libc.so.1`_pread+0xa mysqld`my_pread+0x54 mysqld`_mi_read_static_record+0x67 mysqld`mi_rnext+0x1fe ... mysqld`handle_one_connection+0x855 libc.so.1`_thr_setup+0x67 1564811 .... Replace high cost pread with “--myisam_use_mmap=1”- useful for mysql DBA  improvement ! A 94% performance Dtrace: Memory Analysis  Check which process causes anonymous page in # dtrace -n anonpgin '{@[execname] = count()}' dtrace: description anonpgin matched 1 probe sshd 2 vmstat 23 mysqld 673 Use Dtrace to measure waiting for paging in # ./ whospaging.d – available at Solaris Internals(http://www.solarisinternals.com/si/dtrace/)‫‏‬ Who's on cpu (milliseconds): sshd 1 vmstat 3 mysqld 120 sched 43210 Who's waiting for pagin (milliseconds): mysqld 239082 Dtrace Probes In MySQL  Provide deep view of internal MySQL core server and storage engines' operation & behaviour >Database information >Query execution latency >Index & table scan cost >Wait events inside MyISAM & Innodb >Deadlock information >Query cache hit/miss >And many more... Speed resolution of performance bottlenecks with in database design and MySQL server configurations Negligible performance overhead Easy steps to create & insert your own Dtrace probes into MySQL MySQL DTrace GUI Monitor Tool - Chime     Query execution Time  Enable‫‏‬Slow‫‏‬query‫‏‬log‫‏‬with‫–“‏‬log-slow-queries”‫‏‬requires‫‏‬re-start MySQL server  SQL statements with query execution time longer than “long_query_time”‫‏‬second in the log file  SQL statements generating most loads on the application may not in slow query log  Replication query statements are not available in slow query log  Time spending by the query optimizer to generate query plan is not available in slow query log  Using DTrace can get mising query execution information online Insert DTrace Probes Into MySQL  Step 1: Figure out what probes are needed to insert into the source code  Step 2: Define MySQL Provider and probes # cat mysql_dtrace.d provider mysql { probe query__execute__start(void *, char *, char *, const char *, char *); probe query__execute__finish(void *, char *, char *, const char *, char *,int); } >Two Probes defined in the mysql provider >Note to use two underscore(__) translated to hypen automatically  Step 3:‫‏‬Define‫‏‬a‫‏‬header‫‏‬file‫“‏‬mysql_dtrace.h”‫‏‬with‫‏‬definitions‫‏‬for‫‏‬ probes dtrace -h -s mysql_dtrace.d #ifndef _MYSQL_H #define _MYSQL_H #define DTRACE_QUERY_EXECUTE_START(arg0, arg1, arg2, arg3, arg4) \ __dtrace_mysql___query__execute__start(arg0, arg1, arg2, arg3, arg4)‫‏‬ #define DTRACE_QUERY_EXECUTE_START_ENABLED() \ __dtraceenabled_mysql___query__execute__start()‫‏‬ extern void __dtrace_mysql___query__execute__start(void *, char *, char *, char *, char *); extern int __dtraceenabled_mysql___query__execute__start(void); #endif  Step 4: Insert the probes into source code #include ... bool mysql_execute_command(THD *thd)‫‏‬ { DTRACE_QUERY_EXECUTE_START((void *)thd, thd->db, thd->security_ctx->user, (char *)thd->security_ctx->host_or_ip,thd->query); ... DTRACE_QUERY_EXECUTE_FINISH((void*)thd, thd->db, thd->security_ctx->user, (char *)thd->security_ctx->host_or_ip, thd->query, res ==0 ? 0: -1); go to end; ... }  Step 5: Build MySQL with DTrace  In the Makefile.in, compile 64-bit MySQL with Dtrace mysqld_OBJECTS = $(am_mysqld_OBJECTS) mysql_dtrace.$(OBJEXT)‫‏‬ mysql_dtrace.o:$(top_srcdir)/include/mysql_dtrace.d $(am_mysqld_OBJECTS)‫‏‬ dtrace -G 64 -s $(top_srcdir)/include/mysql_dtrace.d $(am_mysqld_OBJECTS)  Inserting DTrace probes comleted, DTrace probes are ready to use!  Step 6: Use inserted DTrace probes to measure query execution time with other database information #!/usr/sbin/dtrace -s … mysql*:::query-execute-start { self->start = timestamp; } mysql*:::query-execute-finish /self->start/ { this->query = copyinstr(arg4); } mysql*:::query-execute-finish /self->start/ { this->elapsed = (timestamp - self->start) / 1000000; this->who = strjoin(copyinstr(arg2), strjoin("@", copyinstr(arg3))); printf(" %-16.16s %-18.18s %5d %3d %-32.32s\n", arg1 ? copyinstr(arg1)‫‏‬ : ".", this->who, this->elapsed, (int)arg5, this->query); self->start = 0; }  # ./mysqld_qestat.d DATABASE sbtest USER@HOST root@localhost ms RET QUERY 0 0 show tables sbtest sbtest  root@localhost root@localhost 0 0 178 0 show databases select * from sbtest Use the same steps to insert Dtrace probes to /sql/sql_select.cc at the start and end of choose_plan() function to measure the time spent in query optimization optimizer_prune_level=1 reduce query compilation time Reduce optimizer_search_depth or optimizer_search_depth=0   Index & Table scan cost  Identify the places in MySQL source to handle scanning index, and table >Index-scan functions: index_next, index_next_same, index_prev, index_first, index_last >Table-scan functions: rnd_init, rnd_end, rnd_next, rnd_pos >Insert DTrace Probes at the start and before return from the functions can measure the time spending on scanning table or index. mysql*:::innodb-index-next-start { @indexnext[args[0]] = count(); self->inext = timestamp; } mysql*:::innodb-index-next-finish /self->inext/ { @indexnexttime[args[0]] = sum(timestamp - self->inext); self->inext = 0; }  Expensive index-scan/table-scan report from Dtrace requires to optimize schema accordingly Buffer wait in Innodb  Innodb buffer wait is common in I/O-bound MySQL system while reading page synchronous from disk Insert Dtrace probes at: innobase/buf/buf0rea.c, ulint buf_read_page( ulint space, ulint offset)‫‏‬ { ... DTRACE_INNODB_BUFFER_WAIT_START(); /* We do the i/o in the synchronous aio mode to save thread switches: hence TRUE */ count2 = buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, space, tablespace_version, offset); DTRACE_INNODB_BUFFER_WAIT_FINISH(); ... }  Increase innodb_buffer_pool size  Tune SQL to access rows with fewer block reads(i.e. By adding indexes)‫‏‬ Query cache probes  Use DTrace to measure query cache hit and query cache miss to determine how well the query cache is performing > Insert DTrace Probes at: sql/sql_cache.cc function: send_result_to_client > DTrace Test script to report query cache hit and miss counts by query string, and totals mysql*:::query-cache-hit, mysql*:::query-cache-miss { this->query = copyinstr(arg4); } mysql*:::query-cache-hit { @elapsed[this->query, "hit"] = count(); hits++; } mysql*:::query-cache-miss { @elapsed[this->query, "miss"] = count(); misses++; } # ./mysqld_qchit.d Tracing... Hit Ctrl-C to end. ^C QUERY select * from months select * from months where num > 3 select * from months where num > 3 and num < 9 show databases show tables select * from months RESULT COUNT miss hit miss miss miss hit 1 1 1 1 1 9 Hits : 10 Misses : 4 Hit Rate : 71%  Tuning‫“‏‬query_cache_size”‫‏‬variable‫‏‬according‫‏‬to‫‏‬the‫‏‬hit/miss‫‏‬rate DTrace Performance Impact  Inserting DTrace Probes into MySQL source code are useful for MySQL DBA, MySQL & application developers. The performance impact of adding DTrace probes is critical for enterprise environment.  Cost of inserting USDT probe can be basically negligible: Each probe inserted into the source code can be enabled by adding the code like: if (PROVIDER_PROBE_ENABLED()‫‏‬ { PROVIDER_PROBE(arg0,...); } DTrace Probes In MySQL 6.0 probe insert_row_start(); probe insert_row_end(); probe filesort_start(); probe filesort_end(); probe delete_start(); probe delete_end(); probe select_start() probe select_end(); probe update_start(); probe update_end();  compile‫‏‬with‫--“‏‬enable-dtrace”‫‏‬configure‫‏‬option‫‏‬to‫‏‬use‫‏‬ the Dtrace probes Integrate with Chime Tool  Chime is graphical tool for visualizing DTrace aggregations. It provides alternative CLI-based tool output with more visually apprealing and more useful to display data over time  Available to download: http://www.opensolaris.org/os/project/dtrace-chime. One-step installation: > Run pkgadd -d osol0chime--1.4.pkg  chime requires Solaris Nevada build 35 or later > Run local: /opt/OSOL0chime/bin/chime > Run remotely: /opt/OSOL0chime/bin/chime  chime provides wizard to automatically generate new display for DTrace script  Resources > http://www.opensolaris.org/os/community/dtrace - OpenSolaris Community: Dtrace > http://developers.sun.com/solaris/articles/solaris_perftools.html - Solaris Performance and Tools > http://docs.sun.com/app/docs/doc/817-6223/6mlkidlms?a=view - Statically Defined Tracing for User Applications chapter of DTrace Manual > http://www.brendangregg.com - DTrace toolkits > http://www.solarisinternals.com/wiki/index.php/DTrace_Topics_Databases - DTrace Topics Databases  Acknowledgements Brandan Gregg – Sun Microsystems Engineer in Advanced Products Group