Re: pg_upgrade with parallel tablespace copying

Slightly modified patch applied.  This is my last planned pg_upgrade
change for 9.3.

---------------------------------------------------------------------------

On Mon, Jan  7, 2013 at 10:51:21PM -0500, Bruce Momjian wrote:
> Pg_upgrade by default (without --link) copies heap/index files from the
> old to new cluster.  This patch implements parallel heap/index file
> copying in pg_upgrade using the --jobs option.  It uses the same
> infrastructure used for pg_upgrade parallel dump/restore.  Here are the
> performance results:
> 
>                --- seconds ---
>        GB        git    patched
>         2       62.09    63.75
>         4       95.93   107.22
>         8      194.96   195.29
>        16      494.38   348.93
>        32      983.28   644.23
>        64     2227.73  1244.08
>       128     4735.83  2547.09
> 
> Because of the kernel cache, you only see a big win when the amount of
> copy data exceeds the kernel cache.  For testing, I used a 24GB, 16-core
> machine with two magnetic disks with one tablespace on each.  Using more
> tablespaces would yield larger improvements.  My test script is
> attached.  
> 
> I consider this patch ready for application.  This is the last
> pg_upgrade performance improvement idea I am considering.
> 
> -- 
>   Bruce Momjian  <bruce@momjian.us>        http://momjian.us
>   EnterpriseDB                             http://enterprisedb.com
> 
>   + It's impossible for everything to be true. +

> diff --git a/contrib/pg_upgrade/check.c b/contrib/pg_upgrade/check.c
> new file mode 100644
> index 59f8fd0..1780788
> *** a/contrib/pg_upgrade/check.c
> --- b/contrib/pg_upgrade/check.c
> *************** create_script_for_old_cluster_deletion(c
> *** 606,612 ****
>       fprintf(script, RMDIR_CMD " %s\n", fix_path_separator(old_cluster.pgdata));
>   
>       /* delete old cluster's alternate tablespaces */
> !     for (tblnum = 0; tblnum < os_info.num_tablespaces; tblnum++)
>       {
>           /*
>            * Do the old cluster's per-database directories share a directory
> --- 606,612 ----
>       fprintf(script, RMDIR_CMD " %s\n", fix_path_separator(old_cluster.pgdata));
>   
>       /* delete old cluster's alternate tablespaces */
> !     for (tblnum = 0; tblnum < os_info.num_old_tablespaces; tblnum++)
>       {
>           /*
>            * Do the old cluster's per-database directories share a directory
> *************** create_script_for_old_cluster_deletion(c
> *** 621,634 ****
>               /* remove PG_VERSION? */
>               if (GET_MAJOR_VERSION(old_cluster.major_version) <= 804)
>                   fprintf(script, RM_CMD " %s%s%cPG_VERSION\n",
> !                         fix_path_separator(os_info.tablespaces[tblnum]), 
>                           fix_path_separator(old_cluster.tablespace_suffix),
>                           PATH_SEPARATOR);
>   
>               for (dbnum = 0; dbnum < old_cluster.dbarr.ndbs; dbnum++)
>               {
>                   fprintf(script, RMDIR_CMD " %s%s%c%d\n",
> !                         fix_path_separator(os_info.tablespaces[tblnum]),
>                           fix_path_separator(old_cluster.tablespace_suffix),
>                           PATH_SEPARATOR, old_cluster.dbarr.dbs[dbnum].db_oid);
>               }
> --- 621,634 ----
>               /* remove PG_VERSION? */
>               if (GET_MAJOR_VERSION(old_cluster.major_version) <= 804)
>                   fprintf(script, RM_CMD " %s%s%cPG_VERSION\n",
> !                         fix_path_separator(os_info.old_tablespaces[tblnum]), 
>                           fix_path_separator(old_cluster.tablespace_suffix),
>                           PATH_SEPARATOR);
>   
>               for (dbnum = 0; dbnum < old_cluster.dbarr.ndbs; dbnum++)
>               {
>                   fprintf(script, RMDIR_CMD " %s%s%c%d\n",
> !                         fix_path_separator(os_info.old_tablespaces[tblnum]),
>                           fix_path_separator(old_cluster.tablespace_suffix),
>                           PATH_SEPARATOR, old_cluster.dbarr.dbs[dbnum].db_oid);
>               }
> *************** create_script_for_old_cluster_deletion(c
> *** 640,646 ****
>                * or a version-specific subdirectory.
>                */
>               fprintf(script, RMDIR_CMD " %s%s\n",
> !                     fix_path_separator(os_info.tablespaces[tblnum]), 
>                       fix_path_separator(old_cluster.tablespace_suffix));
>       }
>   
> --- 640,646 ----
>                * or a version-specific subdirectory.
>                */
>               fprintf(script, RMDIR_CMD " %s%s\n",
> !                     fix_path_separator(os_info.old_tablespaces[tblnum]), 
>                       fix_path_separator(old_cluster.tablespace_suffix));
>       }
>   
> diff --git a/contrib/pg_upgrade/info.c b/contrib/pg_upgrade/info.c
> new file mode 100644
> index 0c11ff8..7fd4584
> *** a/contrib/pg_upgrade/info.c
> --- b/contrib/pg_upgrade/info.c
> *************** create_rel_filename_map(const char *old_
> *** 106,125 ****
>            * relation belongs to the default tablespace, hence relfiles should
>            * exist in the data directories.
>            */
> !         snprintf(map->old_dir, sizeof(map->old_dir), "%s/base/%u", old_data,
> !                  old_db->db_oid);
> !         snprintf(map->new_dir, sizeof(map->new_dir), "%s/base/%u", new_data,
> !                  new_db->db_oid);
>       }
>       else
>       {
>           /* relation belongs to a tablespace, so use the tablespace location */
> !         snprintf(map->old_dir, sizeof(map->old_dir), "%s%s/%u", old_rel->tablespace,
> !                  old_cluster.tablespace_suffix, old_db->db_oid);
> !         snprintf(map->new_dir, sizeof(map->new_dir), "%s%s/%u", new_rel->tablespace,
> !                  new_cluster.tablespace_suffix, new_db->db_oid);
>       }
>   
>       /*
>        * old_relfilenode might differ from pg_class.oid (and hence
>        * new_relfilenode) because of CLUSTER, REINDEX, or VACUUM FULL.
> --- 106,130 ----
>            * relation belongs to the default tablespace, hence relfiles should
>            * exist in the data directories.
>            */
> !         strlcpy(map->old_tablespace, old_data, sizeof(map->old_tablespace));
> !         strlcpy(map->new_tablespace, new_data, sizeof(map->new_tablespace));
> !         strlcpy(map->old_tablespace_suffix, "/base", sizeof(map->old_tablespace_suffix));
> !         strlcpy(map->new_tablespace_suffix, "/base", sizeof(map->new_tablespace_suffix));
>       }
>       else
>       {
>           /* relation belongs to a tablespace, so use the tablespace location */
> !         strlcpy(map->old_tablespace, old_rel->tablespace, sizeof(map->old_tablespace));
> !         strlcpy(map->new_tablespace, new_rel->tablespace, sizeof(map->new_tablespace));
> !         strlcpy(map->old_tablespace_suffix, old_cluster.tablespace_suffix,
> !                 sizeof(map->old_tablespace_suffix));
> !         strlcpy(map->new_tablespace_suffix, new_cluster.tablespace_suffix,
> !                 sizeof(map->new_tablespace_suffix));
>       }
>   
> +     map->old_db_oid = old_db->db_oid;
> +     map->new_db_oid = new_db->db_oid;
> + 
>       /*
>        * old_relfilenode might differ from pg_class.oid (and hence
>        * new_relfilenode) because of CLUSTER, REINDEX, or VACUUM FULL.
> diff --git a/contrib/pg_upgrade/parallel.c b/contrib/pg_upgrade/parallel.c
> new file mode 100644
> index 8ea36bc..d157511
> *** a/contrib/pg_upgrade/parallel.c
> --- b/contrib/pg_upgrade/parallel.c
> *************** typedef struct {
> *** 34,44 ****
>       char log_file[MAXPGPATH];
>       char opt_log_file[MAXPGPATH];
>       char cmd[MAX_STRING];
> ! } thread_arg;
>   
> ! thread_arg **thread_args;
>   
> ! DWORD win32_exec_prog(thread_arg *args);
>   
>   #endif
>   
> --- 34,57 ----
>       char log_file[MAXPGPATH];
>       char opt_log_file[MAXPGPATH];
>       char cmd[MAX_STRING];
> ! } exec_thread_arg;
>   
> ! typedef struct {
> !     DbInfoArr *old_db_arr;
> !     DbInfoArr *new_db_arr;
> !     char old_pgdata[MAXPGPATH];
> !     char new_pgdata[MAXPGPATH];
> !     char old_tablespace[MAXPGPATH];
> ! } transfer_thread_arg;
>   
> ! exec_thread_arg **exec_thread_args;
> ! transfer_thread_arg **transfer_thread_args;
> ! 
> ! /* track current thread_args struct so reap_child() can be used for all cases */
> ! void **cur_thread_args;
> ! 
> ! DWORD win32_exec_prog(exec_thread_arg *args);
> ! DWORD win32_transfer_all_new_dbs(transfer_thread_arg *args);
>   
>   #endif
>   
> *************** parallel_exec_prog(const char *log_file,
> *** 58,64 ****
>       pid_t        child;
>   #else
>       HANDLE        child;
> !     thread_arg    *new_arg;
>   #endif
>   
>       va_start(args, fmt);
> --- 71,77 ----
>       pid_t        child;
>   #else
>       HANDLE        child;
> !     exec_thread_arg    *new_arg;
>   #endif
>   
>       va_start(args, fmt);
> *************** parallel_exec_prog(const char *log_file,
> *** 71,77 ****
>       else
>       {
>           /* parallel */
> ! 
>           /* harvest any dead children */
>           while (reap_child(false) == true)
>               ;
> --- 84,92 ----
>       else
>       {
>           /* parallel */
> ! #ifdef WIN32
> !         cur_thread_args = (void **)exec_thread_args;
> ! #endif    
>           /* harvest any dead children */
>           while (reap_child(false) == true)
>               ;
> *************** parallel_exec_prog(const char *log_file,
> *** 100,106 ****
>               int i;
>   
>               thread_handles = pg_malloc(user_opts.jobs * sizeof(HANDLE));
> !             thread_args = pg_malloc(user_opts.jobs * sizeof(thread_arg *));
>   
>               /*
>                *    For safety and performance, we keep the args allocated during
> --- 115,121 ----
>               int i;
>   
>               thread_handles = pg_malloc(user_opts.jobs * sizeof(HANDLE));
> !             exec_thread_args = pg_malloc(user_opts.jobs * sizeof(exec_thread_arg *));
>   
>               /*
>                *    For safety and performance, we keep the args allocated during
> *************** parallel_exec_prog(const char *log_file,
> *** 108,118 ****
>                *    in a thread different from the one that allocated it.
>                */
>               for (i = 0; i < user_opts.jobs; i++)
> !                 thread_args[i] = pg_malloc(sizeof(thread_arg));
>           }
>   
>           /* use first empty array element */
> !         new_arg = thread_args[parallel_jobs-1];
>   
>           /* Can only pass one pointer into the function, so use a struct */
>           strcpy(new_arg->log_file, log_file);
> --- 123,133 ----
>                *    in a thread different from the one that allocated it.
>                */
>               for (i = 0; i < user_opts.jobs; i++)
> !                 exec_thread_args[i] = pg_malloc(sizeof(exec_thread_arg));
>           }
>   
>           /* use first empty array element */
> !         new_arg = exec_thread_args[parallel_jobs-1];
>   
>           /* Can only pass one pointer into the function, so use a struct */
>           strcpy(new_arg->log_file, log_file);
> *************** parallel_exec_prog(const char *log_file,
> *** 134,140 ****
>   
>   #ifdef WIN32
>   DWORD
> ! win32_exec_prog(thread_arg *args)
>   {
>       int ret;
>   
> --- 149,155 ----
>   
>   #ifdef WIN32
>   DWORD
> ! win32_exec_prog(exec_thread_arg *args)
>   {
>       int ret;
>   
> *************** win32_exec_prog(thread_arg *args)
> *** 147,152 ****
> --- 162,273 ----
>   
>   
>   /*
> +  *    parallel_transfer_all_new_dbs
> +  *
> +  *    This has the same API as transfer_all_new_dbs, except it does parallel execution
> +  *    by transfering multiple tablespaces in parallel
> +  */
> + void parallel_transfer_all_new_dbs(DbInfoArr *old_db_arr, DbInfoArr *new_db_arr,
> +                                    char *old_pgdata, char *new_pgdata,
> +                                    char *old_tablespace)
> + {
> + #ifndef WIN32
> +     pid_t        child;
> + #else
> +     HANDLE        child;
> +     transfer_thread_arg    *new_arg;
> + #endif
> + 
> +     if (user_opts.jobs <= 1)
> +         /* throw_error must be true to allow jobs */
> +         transfer_all_new_dbs(old_db_arr, new_db_arr, old_pgdata, new_pgdata, NULL);
> +     else
> +     {
> +         /* parallel */
> + #ifdef WIN32
> +         cur_thread_args = (void **)transfer_thread_args;
> + #endif
> +         /* harvest any dead children */
> +         while (reap_child(false) == true)
> +             ;
> + 
> +         /* must we wait for a dead child? */
> +         if (parallel_jobs >= user_opts.jobs)
> +             reap_child(true);
> +             
> +         /* set this before we start the job */
> +         parallel_jobs++;
> +     
> +         /* Ensure stdio state is quiesced before forking */
> +         fflush(NULL);
> + 
> + #ifndef WIN32
> +         child = fork();
> +         if (child == 0)
> +         {
> +             transfer_all_new_dbs(old_db_arr, new_db_arr, old_pgdata, new_pgdata,
> +                                  old_tablespace);
> +             /* if we take another exit path, it will be non-zero */
> +             /* use _exit to skip atexit() functions */
> +             _exit(0);
> +         }
> +         else if (child < 0)
> +             /* fork failed */
> +             pg_log(PG_FATAL, "could not create worker process: %s\n", strerror(errno));
> + #else
> +         if (thread_handles == NULL)
> +         {
> +             int i;
> + 
> +             thread_handles = pg_malloc(user_opts.jobs * sizeof(HANDLE));
> +             transfer_thread_args = pg_malloc(user_opts.jobs * sizeof(transfer_thread_arg *));
> + 
> +             /*
> +              *    For safety and performance, we keep the args allocated during
> +              *    the entire life of the process, and we don't free the args
> +              *    in a thread different from the one that allocated it.
> +              */
> +             for (i = 0; i < user_opts.jobs; i++)
> +                 transfer_thread_args[i] = pg_malloc(sizeof(transfer_thread_arg));
> +         }
> + 
> +         /* use first empty array element */
> +         new_arg = transfer_thread_args[parallel_jobs-1];
> + 
> +         /* Can only pass one pointer into the function, so use a struct */
> +         new_arg->old_db_arr = old_db_arr;
> +         new_arg->new_db_arr = new_db_arr;
> +         strcpy(new_arg->old_pgdata, old_pgdata);
> +         strcpy(new_arg->new_pgdata, new_pgdata);
> +         strcpy(new_arg->old_tablespace, old_tablespace);
> + 
> +         child = (HANDLE) _beginthreadex(NULL, 0, (void *) win32_exec_prog,
> +                         new_arg, 0, NULL);
> +         if (child == 0)
> +             pg_log(PG_FATAL, "could not create worker thread: %s\n", strerror(errno));
> + 
> +         thread_handles[parallel_jobs-1] = child;
> + #endif
> +     }
> + 
> +     return;
> + }
> + 
> + 
> + #ifdef WIN32
> + DWORD
> + win32_transfer_all_new_dbs(transfer_thread_arg *args)
> + {
> +     transfer_all_new_dbs(args->old_db_arr, args->new_db_arr, args->old_pgdata,
> +                          args->new_pgdata, args->old_tablespace);
> + 
> +     /* terminates thread */
> +     return 0;
> + }
> + #endif
> + 
> + 
> + /*
>    *    collect status from a completed worker child
>    */
>   bool
> *************** reap_child(bool wait_for_child)
> *** 195,201 ****
>       /*    Move last slot into dead child's position */
>       if (thread_num != parallel_jobs - 1)
>       {
> !         thread_arg *tmp_args;
>       
>           thread_handles[thread_num] = thread_handles[parallel_jobs - 1];
>   
> --- 316,322 ----
>       /*    Move last slot into dead child's position */
>       if (thread_num != parallel_jobs - 1)
>       {
> !         void *tmp_args;
>       
>           thread_handles[thread_num] = thread_handles[parallel_jobs - 1];
>   
> *************** reap_child(bool wait_for_child)
> *** 205,213 ****
>            *    reused by the next created thread.  Instead, the new thread
>            *    will use the arg struct of the thread that just died.
>            */
> !         tmp_args = thread_args[thread_num];
> !         thread_args[thread_num] = thread_args[parallel_jobs - 1];
> !         thread_args[parallel_jobs - 1] = tmp_args;
>       }
>   #endif
>   
> --- 326,334 ----
>            *    reused by the next created thread.  Instead, the new thread
>            *    will use the arg struct of the thread that just died.
>            */
> !         tmp_args = cur_thread_args[thread_num];
> !         cur_thread_args[thread_num] = cur_thread_args[parallel_jobs - 1];
> !         cur_thread_args[parallel_jobs - 1] = tmp_args;
>       }
>   #endif
>   
> diff --git a/contrib/pg_upgrade/pg_upgrade.c b/contrib/pg_upgrade/pg_upgrade.c
> new file mode 100644
> index 70c749d..85997e5
> *** a/contrib/pg_upgrade/pg_upgrade.c
> --- b/contrib/pg_upgrade/pg_upgrade.c
> *************** main(int argc, char **argv)
> *** 133,139 ****
>       if (user_opts.transfer_mode == TRANSFER_MODE_LINK)
>           disable_old_cluster();
>   
> !     transfer_all_new_dbs(&old_cluster.dbarr, &new_cluster.dbarr,
>                            old_cluster.pgdata, new_cluster.pgdata);
>   
>       /*
> --- 133,139 ----
>       if (user_opts.transfer_mode == TRANSFER_MODE_LINK)
>           disable_old_cluster();
>   
> !     transfer_all_new_tablespaces(&old_cluster.dbarr, &new_cluster.dbarr,
>                            old_cluster.pgdata, new_cluster.pgdata);
>   
>       /*
> diff --git a/contrib/pg_upgrade/pg_upgrade.h b/contrib/pg_upgrade/pg_upgrade.h
> new file mode 100644
> index c1a2f53..d5c3fa9
> *** a/contrib/pg_upgrade/pg_upgrade.h
> --- b/contrib/pg_upgrade/pg_upgrade.h
> *************** typedef struct
> *** 134,141 ****
>    */
>   typedef struct
>   {
> !     char        old_dir[MAXPGPATH];
> !     char        new_dir[MAXPGPATH];
>   
>       /*
>        * old/new relfilenodes might differ for pg_largeobject(_metadata) indexes
> --- 134,145 ----
>    */
>   typedef struct
>   {
> !     char        old_tablespace[MAXPGPATH];
> !     char        new_tablespace[MAXPGPATH];
> !     char        old_tablespace_suffix[MAXPGPATH];
> !     char        new_tablespace_suffix[MAXPGPATH];
> !     Oid            old_db_oid;
> !     Oid            new_db_oid;
>   
>       /*
>        * old/new relfilenodes might differ for pg_largeobject(_metadata) indexes
> *************** typedef struct
> *** 276,283 ****
>       const char *progname;        /* complete pathname for this program */
>       char       *exec_path;        /* full path to my executable */
>       char       *user;            /* username for clusters */
> !     char      **tablespaces;    /* tablespaces */
> !     int            num_tablespaces;
>       char      **libraries;        /* loadable libraries */
>       int            num_libraries;
>       ClusterInfo *running_cluster;
> --- 280,287 ----
>       const char *progname;        /* complete pathname for this program */
>       char       *exec_path;        /* full path to my executable */
>       char       *user;            /* username for clusters */
> !     char      **old_tablespaces;    /* tablespaces */
> !     int            num_old_tablespaces;
>       char      **libraries;        /* loadable libraries */
>       int            num_libraries;
>       ClusterInfo *running_cluster;
> *************** void        get_sock_dir(ClusterInfo *cluster,
> *** 398,406 ****
>   /* relfilenode.c */
>   
>   void        get_pg_database_relfilenode(ClusterInfo *cluster);
> ! void        transfer_all_new_dbs(DbInfoArr *olddb_arr,
> !                    DbInfoArr *newdb_arr, char *old_pgdata, char *new_pgdata);
> ! 
>   
>   /* tablespace.c */
>   
> --- 402,412 ----
>   /* relfilenode.c */
>   
>   void        get_pg_database_relfilenode(ClusterInfo *cluster);
> ! void        transfer_all_new_tablespaces(DbInfoArr *old_db_arr,
> !                    DbInfoArr *new_db_arr, char *old_pgdata, char *new_pgdata);
> ! void        transfer_all_new_dbs(DbInfoArr *old_db_arr,
> !                    DbInfoArr *new_db_arr, char *old_pgdata, char *new_pgdata,
> !                    char *old_tablespace);
>   
>   /* tablespace.c */
>   
> *************** void old_8_3_invalidate_bpchar_pattern_o
> *** 464,472 ****
>   char       *old_8_3_create_sequence_script(ClusterInfo *cluster);
>   
>   /* parallel.c */
> ! void parallel_exec_prog(const char *log_file, const char *opt_log_file,
>             const char *fmt,...)
>   __attribute__((format(PG_PRINTF_ATTRIBUTE, 3, 4)));
> ! 
> ! bool reap_child(bool wait_for_child);
>   
> --- 470,480 ----
>   char       *old_8_3_create_sequence_script(ClusterInfo *cluster);
>   
>   /* parallel.c */
> ! void        parallel_exec_prog(const char *log_file, const char *opt_log_file,
>             const char *fmt,...)
>   __attribute__((format(PG_PRINTF_ATTRIBUTE, 3, 4)));
> ! void        parallel_transfer_all_new_dbs(DbInfoArr *old_db_arr, DbInfoArr *new_db_arr,
> !                                           char *old_pgdata, char *new_pgdata,
> !                                           char *old_tablespace);
> ! bool        reap_child(bool wait_for_child);
>   
> diff --git a/contrib/pg_upgrade/relfilenode.c b/contrib/pg_upgrade/relfilenode.c
> new file mode 100644
> index 9d0d5a0..cfdd39c
> *** a/contrib/pg_upgrade/relfilenode.c
> --- b/contrib/pg_upgrade/relfilenode.c
> ***************
> *** 16,42 ****
>   
>   
>   static void transfer_single_new_db(pageCnvCtx *pageConverter,
> !                        FileNameMap *maps, int size);
>   static void transfer_relfile(pageCnvCtx *pageConverter, FileNameMap *map,
>                                const char *suffix);
>   
>   
>   /*
>    * transfer_all_new_dbs()
>    *
>    * Responsible for upgrading all database. invokes routines to generate mappings and then
>    * physically link the databases.
>    */
>   void
> ! transfer_all_new_dbs(DbInfoArr *old_db_arr,
> !                    DbInfoArr *new_db_arr, char *old_pgdata, char *new_pgdata)
>   {
>       int            old_dbnum,
>                   new_dbnum;
>   
> -     pg_log(PG_REPORT, "%s user relation files\n",
> -       user_opts.transfer_mode == TRANSFER_MODE_LINK ? "Linking" : "Copying");
> - 
>       /* Scan the old cluster databases and transfer their files */
>       for (old_dbnum = new_dbnum = 0;
>            old_dbnum < old_db_arr->ndbs;
> --- 16,81 ----
>   
>   
>   static void transfer_single_new_db(pageCnvCtx *pageConverter,
> !                        FileNameMap *maps, int size, char *old_tablespace);
>   static void transfer_relfile(pageCnvCtx *pageConverter, FileNameMap *map,
>                                const char *suffix);
>   
>   
>   /*
> +  * transfer_all_new_tablespaces()
> +  *
> +  * Responsible for upgrading all database. invokes routines to generate mappings and then
> +  * physically link the databases.
> +  */
> + void
> + transfer_all_new_tablespaces(DbInfoArr *old_db_arr, DbInfoArr *new_db_arr,
> +                     char *old_pgdata, char *new_pgdata)
> + {
> +     pg_log(PG_REPORT, "%s user relation files\n",
> +       user_opts.transfer_mode == TRANSFER_MODE_LINK ? "Linking" : "Copying");
> + 
> +     /*
> +      *    Transfering files by tablespace is tricky because a single database
> +      *    can use multiple tablespaces.  For non-parallel mode, we just pass a
> +      *    NULL tablespace path, which matches all tablespaces.  In parallel mode,
> +      *    we pass the default tablespace and all user-created tablespaces
> +      *    and let those operations happen in parallel.
> +      */
> +     if (user_opts.jobs <= 1)
> +         parallel_transfer_all_new_dbs(old_db_arr, new_db_arr, old_pgdata,
> +                                       new_pgdata, NULL);
> +     else
> +     {
> +         int tblnum;
> +         
> +         for (tblnum = 0; tblnum < os_info.num_old_tablespaces; tblnum++)
> +             parallel_transfer_all_new_dbs(old_db_arr, new_db_arr, old_pgdata,
> +                                   new_pgdata, os_info.old_tablespaces[tblnum]);
> +         /* reap all children */
> +         while (reap_child(true) == true)
> +             ;
> +     }
> + 
> +     end_progress_output();
> +     check_ok();
> + 
> +     return;
> + }
> + 
> + 
> + /*
>    * transfer_all_new_dbs()
>    *
>    * Responsible for upgrading all database. invokes routines to generate mappings and then
>    * physically link the databases.
>    */
>   void
> ! transfer_all_new_dbs(DbInfoArr *old_db_arr, DbInfoArr *new_db_arr,
> !                     char *old_pgdata, char *new_pgdata, char *old_tablespace)
>   {
>       int            old_dbnum,
>                   new_dbnum;
>   
>       /* Scan the old cluster databases and transfer their files */
>       for (old_dbnum = new_dbnum = 0;
>            old_dbnum < old_db_arr->ndbs;
> *************** transfer_all_new_dbs(DbInfoArr *old_db_a
> *** 75,89 ****
>   #ifdef PAGE_CONVERSION
>               pageConverter = setupPageConverter();
>   #endif
> !             transfer_single_new_db(pageConverter, mappings, n_maps);
>   
>               pg_free(mappings);
>           }
>       }
>   
> -     end_progress_output();
> -     check_ok();
> - 
>       return;
>   }
>   
> --- 114,126 ----
>   #ifdef PAGE_CONVERSION
>               pageConverter = setupPageConverter();
>   #endif
> !             transfer_single_new_db(pageConverter, mappings, n_maps,
> !                                    old_tablespace);
>   
>               pg_free(mappings);
>           }
>       }
>   
>       return;
>   }
>   
> *************** get_pg_database_relfilenode(ClusterInfo
> *** 125,131 ****
>    */
>   static void
>   transfer_single_new_db(pageCnvCtx *pageConverter,
> !                        FileNameMap *maps, int size)
>   {
>       int            mapnum;
>       bool        vm_crashsafe_match = true;
> --- 162,168 ----
>    */
>   static void
>   transfer_single_new_db(pageCnvCtx *pageConverter,
> !                        FileNameMap *maps, int size, char *old_tablespace)
>   {
>       int            mapnum;
>       bool        vm_crashsafe_match = true;
> *************** transfer_single_new_db(pageCnvCtx *pageC
> *** 140,157 ****
>   
>       for (mapnum = 0; mapnum < size; mapnum++)
>       {
> !         /* transfer primary file */
> !         transfer_relfile(pageConverter, &maps[mapnum], "");
> ! 
> !         /* fsm/vm files added in PG 8.4 */
> !         if (GET_MAJOR_VERSION(old_cluster.major_version) >= 804)
>           {
> !             /*
> !              * Copy/link any fsm and vm files, if they exist
> !              */
> !             transfer_relfile(pageConverter, &maps[mapnum], "_fsm");
> !             if (vm_crashsafe_match)
> !                 transfer_relfile(pageConverter, &maps[mapnum], "_vm");
>           }
>       }
>   }
> --- 177,198 ----
>   
>       for (mapnum = 0; mapnum < size; mapnum++)
>       {
> !         if (old_tablespace == NULL ||
> !             strcmp(maps[mapnum].old_tablespace, old_tablespace) == 0)
>           {
> !             /* transfer primary file */
> !             transfer_relfile(pageConverter, &maps[mapnum], "");
> !     
> !             /* fsm/vm files added in PG 8.4 */
> !             if (GET_MAJOR_VERSION(old_cluster.major_version) >= 804)
> !             {
> !                 /*
> !                  * Copy/link any fsm and vm files, if they exist
> !                  */
> !                 transfer_relfile(pageConverter, &maps[mapnum], "_fsm");
> !                 if (vm_crashsafe_match)
> !                     transfer_relfile(pageConverter, &maps[mapnum], "_vm");
> !             }
>           }
>       }
>   }
> *************** transfer_relfile(pageCnvCtx *pageConvert
> *** 187,196 ****
>           else
>               snprintf(extent_suffix, sizeof(extent_suffix), ".%d", segno);
>   
> !         snprintf(old_file, sizeof(old_file), "%s/%u%s%s", map->old_dir,
> !                  map->old_relfilenode, type_suffix, extent_suffix);
> !         snprintf(new_file, sizeof(new_file), "%s/%u%s%s", map->new_dir,
> !                  map->new_relfilenode, type_suffix, extent_suffix);
>       
>           /* Is it an extent, fsm, or vm file? */
>           if (type_suffix[0] != '\0' || segno != 0)
> --- 228,239 ----
>           else
>               snprintf(extent_suffix, sizeof(extent_suffix), ".%d", segno);
>   
> !         snprintf(old_file, sizeof(old_file), "%s%s/%u/%u%s%s", map->old_tablespace,
> !                  map->old_tablespace_suffix, map->old_db_oid, map->old_relfilenode,
> !                  type_suffix, extent_suffix);
> !         snprintf(new_file, sizeof(new_file), "%s%s/%u/%u%s%s", map->new_tablespace,
> !                  map->new_tablespace_suffix, map->new_db_oid, map->new_relfilenode,
> !                  type_suffix, extent_suffix);
>       
>           /* Is it an extent, fsm, or vm file? */
>           if (type_suffix[0] != '\0' || segno != 0)
> *************** transfer_relfile(pageCnvCtx *pageConvert
> *** 239,241 ****
> --- 282,285 ----
>   
>       return;
>   }
> + 
> diff --git a/contrib/pg_upgrade/tablespace.c b/contrib/pg_upgrade/tablespace.c
> new file mode 100644
> index a93c517..321738d
> *** a/contrib/pg_upgrade/tablespace.c
> --- b/contrib/pg_upgrade/tablespace.c
> *************** init_tablespaces(void)
> *** 23,29 ****
>       set_tablespace_directory_suffix(&old_cluster);
>       set_tablespace_directory_suffix(&new_cluster);
>   
> !     if (os_info.num_tablespaces > 0 &&
>       strcmp(old_cluster.tablespace_suffix, new_cluster.tablespace_suffix) == 0)
>           pg_log(PG_FATAL,
>                  "Cannot upgrade to/from the same system catalog version when\n"
> --- 23,29 ----
>       set_tablespace_directory_suffix(&old_cluster);
>       set_tablespace_directory_suffix(&new_cluster);
>   
> !     if (os_info.num_old_tablespaces > 0 &&
>       strcmp(old_cluster.tablespace_suffix, new_cluster.tablespace_suffix) == 0)
>           pg_log(PG_FATAL,
>                  "Cannot upgrade to/from the same system catalog version when\n"
> *************** get_tablespace_paths(void)
> *** 57,72 ****
>   
>       res = executeQueryOrDie(conn, "%s", query);
>   
> !     if ((os_info.num_tablespaces = PQntuples(res)) != 0)
> !         os_info.tablespaces = (char **) pg_malloc(
> !                                    os_info.num_tablespaces * sizeof(char *));
>       else
> !         os_info.tablespaces = NULL;
>   
>       i_spclocation = PQfnumber(res, "spclocation");
>   
> !     for (tblnum = 0; tblnum < os_info.num_tablespaces; tblnum++)
> !         os_info.tablespaces[tblnum] = pg_strdup(
>                                        PQgetvalue(res, tblnum, i_spclocation));
>   
>       PQclear(res);
> --- 57,72 ----
>   
>       res = executeQueryOrDie(conn, "%s", query);
>   
> !     if ((os_info.num_old_tablespaces = PQntuples(res)) != 0)
> !         os_info.old_tablespaces = (char **) pg_malloc(
> !                                    os_info.num_old_tablespaces * sizeof(char *));
>       else
> !         os_info.old_tablespaces = NULL;
>   
>       i_spclocation = PQfnumber(res, "spclocation");
>   
> !     for (tblnum = 0; tblnum < os_info.num_old_tablespaces; tblnum++)
> !         os_info.old_tablespaces[tblnum] = pg_strdup(
>                                        PQgetvalue(res, tblnum, i_spclocation));
>   
>       PQclear(res);
> diff --git a/doc/src/sgml/pgupgrade.sgml b/doc/src/sgml/pgupgrade.sgml
> new file mode 100644
> index 53781e4..3e5d548
> *** a/doc/src/sgml/pgupgrade.sgml
> --- b/doc/src/sgml/pgupgrade.sgml
> *************** NET STOP pgsql-8.3  (<productname>Postgr
> *** 342,351 ****
>   
>       <para>
>        The <option>--jobs</> option allows multiple CPU cores to be used
> !      to dump and reload database schemas in parallel;  a good place to
> !      start is the number of CPU cores on the server.  This option can
> !      dramatically reduce the time to upgrade a multi-database server
> !      running on a multiprocessor machine.
>       </para>
>   
>       <para>
> --- 342,353 ----
>   
>       <para>
>        The <option>--jobs</> option allows multiple CPU cores to be used
> !      for file copy operations and to dump and reload database schemas in
> !      parallel;  a good place to start is the number of CPU cores on the
> !      server, or the number of tablespaces if not using the
> !      <option>--link</> option.  This option can dramatically reduce the
> !      time to upgrade a multi-database server running on a multiprocessor
> !      machine.
>       </para>
>   
>       <para>

> :
> 
> . traprm
> 
> export QUIET=$((QUIET + 1))
> 
> > /rtmp/out
> 
> OLDREL=9.2
> NEWREL=9.3
> BRANCH=tablespace
> SSD="f"
> 
> export PGOPTIONS="-c synchronous_commit=off"
> 
> for CYCLES in 2 4 8 16 32 64 128
> do
>     echo "$CYCLES" >> /rtmp/out
>     for JOBLIMIT in 1 2
>     do
>         cd /pgsql/$REL
>         pgsw $BRANCH
>         cd -
>         tools/setup $OLDREL $NEWREL
>         sleep 2
>         [ "$SSD" = "f" ] && tools/mv_to_archive
>     
>         # need for +16k
>         for CONFIG in /u/pgsql.old/data/postgresql.conf /u/pgsql/data/postgresql.conf
>         do
>             pipe sed 's/#max_locks_per_transaction = 64/max_locks_per_transaction = 64000/' "$CONFIG"
>             pipe sed 's/shared_buffers = 128MB/shared_buffers = 1GB/' "$CONFIG"
>         done
> 
>         pgstart -w /u/pgsql.old/data
>         for DIR in /archive/tmp/t1 /backup0/tmp/t2
>         do    rm -rf "$DIR"
>             mkdir "$DIR"
>             chown postgres "$DIR"
>             chmod 0700 "$DIR"
>         done
> 
>         echo "CREATE TABLESPACE t1 LOCATION '/archive/tmp/t1';" | sql --echo-all test
>         echo "CREATE TABLESPACE t2 LOCATION '/backup0/tmp/t2';" | sql --echo-all test
> 
>         for SPNO in $(jot 2)
>         do
>             for TBLNO in $(jot $(($CYCLES / 2)) )
>             do 
>                 echo "CREATE TABLE test${SPNO}_${TBLNO} (x TEXT) TABLESPACE t$SPNO;"
>                 echo "ALTER TABLE test${SPNO}_${TBLNO} ALTER COLUMN x SET STORAGE EXTERNAL;"
>                 echo "INSERT INTO test${SPNO}_${TBLNO} SELECT repeat('x', 999000000);"
>             done |
>             PGOPTIONS="-c synchronous_commit=off" sql --single-transaction --echo-all test
>         done
>         pgstop /u/pgsql.old/data
>         sleep 2
>         # clear cache
>         echo 3 > /proc/sys/vm/drop_caches
>         # allow system to repopulate
>         sleep 15
>         /usr/bin/time --output=/rtmp/out --append --format '%e' tools/upgrade -j $JOBLIMIT || exit 
>         sleep 2
>     done
> done
> 
> bell

> 
> -- 
> Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-hackers


--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +