Oracle FAQ Your Portal to the Oracle Knowledge Grid
HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US
 

Home -> Community -> Mailing Lists -> Oracle-L -> Ora-04031

Ora-04031

From: Philip Lima <plima_at_sctcorp.com>
Date: Tue, 25 Jul 2000 15:21:56 -0400
Message-Id: <10569.112924@fatcity.com>


This is from Oracle as well, a technical Document on Shared Pool tuning on this error.

Regards,
Phil Lima
SCT Global Government Solutions

                                                                                           
                                                                                           
                                                                                           
                                                                                           
                                                                                           
                                                                                           
                                                                                           
                                                             Subject:                      
                                                             TUNING THE SHARED POOL AND    
                                                             RESOLVING ORA-4031            
                                                             Type:                         
                                                             BULLETIN                      
                                                             Status:                       
                                                             PUBLISHED                     
                                                                                           
                                                                                           
                                                                                           
                                                                                           
                                                             Content Type:                 
                                                              TEXT/PLAIN                   
                                                             Creation Date:                
                                                              12-SEP-1997                  
                                                             Last Revision Date:           
                                                              23-MAY-2000                  
                                                             Language:                     
                                                              USAENG                       
                                                                                           
                                                                                           
                                                             PURPOSE                       
                                                               This document discusses     
                                                             some of the common issues     
                                                             associated with the shared    
                                                               pool in Oracle7 and         
                                                             describes how to diagnose and 
                                                             respond to these issues.      
                                                                                           
                                                             RELATED DOCUMENTS             
                                                               [NOTE:1012049.6]   TUNING   
                                                             LIBRARY CACHE LATCH           
                                                             CONTENTION                    
                                                                                           
                                                                                           
                                                             With Release 7.2 and 7.3,     
                                                             changes have been made to     
                                                             reduce usage of shared        
                                                             memory as well as per-user    
                                                             (UGA) memory.  Also, memory   
                                                             is not being allocated        
                                                             in large contiguous chunks,   
                                                             resulting in better           
                                                             shared-pool utilization and   
                                                             reduction in fragmentation.   
                                                                                           
                                                             1) MEMORY FRAGMENTATION       
                                                                                           
                                                             The primary problem that      
                                                             occurs is that free memory in 
                                                             the shared pool becomes       
                                                             fragmented into small pieces  
                                                             over time.  Any attempt to    
                                                             allocate a large piece        
                                                             of memory in the shared pool  
                                                             will cause large amount of    
                                                             objects in the library        
                                                             cache to be flushed out and   
                                                             may result in an ORA-04031    
                                                             out of shared memory          
                                                             error.                        
                                                                                           
                                                             A) DIAGNOSIS OF FRAGMENTATION 
                                                                                           
                                                             i) ORA-04031 ERROR            
                                                                                           
                                                             One way to diagnose that this 
                                                             is happening is to look for   
                                                             ORA-04031 errors              
                                                             being returned from           
                                                             applications.  When an        
                                                             attempt is made to allocate a 
                                                             large                         
                                                             contiguous piece of shared    
                                                             memory, and not enough        
                                                             contiguous memory can be      
                                                             created in the shared pool,   
                                                             the database will signal this 
                                                             error.                        
                                                                                           
                                                             Before this error is          
                                                             signalled, all objects in the 
                                                             shared pool that are not      
                                                             currently in use will be      
                                                             flushed from the shared pool, 
                                                             and their memory will be      
                                                             freed and merged.  This error 
                                                             only occurs when there is     
                                                             still not a large             
                                                             enough contiguous piece of    
                                                             free memory after this        
                                                             happens.  There may be very   
                                                             large amounts of total free   
                                                             memory in the shared pool,    
                                                             but just not enough           
                                                             contiguous memory.            
                                                                                           
                                                             ii) INIT.ORA PARAMETER        
                                                                                           
                                                             An init.ora parameter can be  
                                                             set so that whenever an       
                                                             ORA-04031 error is            
                                                             signalled a dump will occur   
                                                             into a trace file.  By        
                                                             looking for these trace       
                                                             files, the DBA can determine  
                                                             that these errors are         
                                                             occurring.  This is useful    
                                                             when applications do not      
                                                             always report errors          
                                                             signalled by oracle, or if    
                                                             users                         
                                                             do not report the errors to   
                                                             the DBAs.  The parameter is   
                                                             the following:                
                                                                                           
                                                                 event = "4031 trace name  
                                                             errorstack"                   
                                                                                           
                                                             If you are using 7.0.16 or    
                                                             higher you can use the        
                                                             following:                    
                                                                                           
                                                                 event = "4031 trace name  
                                                             errorstack level 4"           
                                                                                           
                                                             This will cause a dump of the 
                                                             Oracle state objects to occur 
                                                             when this error is            
                                                             signalled.  By looking in the 
                                                             dump for 'load=X' and then    
                                                             looking up a few lines        
                                                             for 'name=' you can often     
                                                             tell whether an object was    
                                                             being loaded into the         
                                                             shared pool when this error   
                                                             occurred.  If an object was   
                                                             being loaded then it is       
                                                             likely that this load is the  
                                                             cause of the problem and the  
                                                             object should be              

'kept' in the shared pool.
The object being loaded is the object printed after the 'name='. Do not use the
'level 4' option in versions
before 7.0.16 because a bug existed that often caused the system to crash with this option enabled due to a latch level violation. Prior to version 7.3, there were a handful of cases where the RDBMS or PL/SQL would attempt to allocate large pieces of contiguous memory. Most of this has been fixed for 7.3. This problem was especially acute when running MTS, when the UGA would be located in the SGA. This should also be fixed in 7.3 and using MTS for a high OLTP scenario is recommended. As a result of all these changes, the ORA-04031 error should be virtually eliminated. If an ORA-04031 error is signalled, quite likely the shared pool is over 90% utilized and the alternative is to increase the shared pool. The only known situation is PL/SQL packages (like STANDARD) where the package contains a very large number (over 400) procedure/function definitions. This still needs to be in contiguous memory and may request memory chunks as large as 15K. Packages like this should be the only ones that should be kept. iii) X$KSMLRU There is a fixed table called x$ksmlru that tracks allocations in the shared pool that cause other objects in the shared pool to be aged out. This fixed table can be used to identify what is causing the large allocation. The columns of this fixed table are the following: KSMLRCOM - allocation comment that describes the type of allocation. If this comment is something like 'MPCODE' or 'PLSQL%' then there is a large PL/SQL object being loaded into the shared pool. This PL/SQL object will need to be 'kept' in the shared pool. If this comment is 'kgltbtab' then the allocation is for a dependency table in the library cache. This is only a problem when several hundred users are logged on using distinct user ids. The solution in this case is to use fully qualified names for all table references. If you are running MTS and the comment is something like
'Fixed UGA' then the
problem is that the init.ora parameter 'open_cursors' is set too high. KSMLRSIZ - amount of contiguous memory being allocated. Values over around 5K start to be a problem, values over 10K are a serious problem, and values over 20K are very serious problems. Anything less then 5K should not be a problem. KSMLRNUM - number of objects that were flushed from the shared pool in order allocate the memory. In release 7.1.3 or later, the following columns also exist: KSMLRHON - the name of the object being loaded into the shared pool if the object is a PL/SQL object or a cursor. KSMLROHV - hash value of object being loaded KSMLRSES - SADDR of the session that loaded the object. The advantage of X$KSMLRU is that it allows you to identify problems with fragmentation that are effecting performance, but that are not bad enough to be causing ORA-04031 errors to be signalled. If a lot of objects are being periodically flushed from the shared pool then this will cause response time problems and will likely cause library cache latch contention problems when the objects are reloaded into the shared pool. With version 7.2, the library cache latch contention should be significantly reduced with the breaking up of the library cache pin latch into a configurable set of symmetric library cache latches. One unusual thing about the x$ksmlru fixed table is that the contents of the fixed table are erased whenever someone selects from the fixed table. This is done since the fixed table stores only the largest allocations that have occurred. The values are reset after being selected so that subsequent large allocations can be noted even if they were not quite as large as others that occurred previously. Because of this resetting, the output of selecting from this table should be carefully noted since it cannot be reselected if it is forgotten. Also you should take care that there are not multiple people on one database that select from this table because only one of them will select the real data. To monitor this fixed table just run the following: select * from x$ksmlru where ksmlrsiz > 5000; iv) MTS Oracle users using SQL*Net V2 can connect to the database using dedicated servers, or multiple clients can use a pool of shared (or MTS) servers. The biggest memory implication of this mode is that the session memory (also known as the UGA) for every session needs to be accessible to every MTS server. This implies that the logical UGA comes out of the physical SGA (or the shared pool) instead of the PGA (process memory). In versions prior to 7.3, there were a few components in the UGA that would request large contiguous chunks of memory, contributing to fragmentation of the shared pool if using MTS. If the system had been up for a while, users would have failures when attempting to connect or executing sql. Starting with 7.3, all these allocations have been segmented such that the average size of memory chunks allocated to the UGA should be about 5K. B) CORRECTION OF FRAGMENTATION i) KEEPING OBJECTS The primary source of problems is large PL/SQL objects. The means of correcting these errors is to 'keep' large PL/SQL object in the shared pool at startup time. This will load the objects into the shared pool and will make sure that the objects are never aged out of the shared pool. If the objects are never aged out then there will not be a problem with trying to load them and not having enough memory. Objects are 'kept' in the shared pool using the dbms_shared_pool package that is defined in the dbmspool.sql file. For example: execute dbms_shared_pool.keep ('SYS.STANDARD'); All large packages that are shipped should be 'kept' if the customer uses PL/SQL. This includes
'STANDARD', 'DBMS_STANDARD',
and 'DIUTIL'. With 7.3, the only package left in this list is 'STANDARD'. All large customer packages should also be marked 'kept'. To mark all packages in the system 'kept' execute the following: declare own varchar2(100); nam varchar2(100); cursor pkgs is select owner, object_name from dba_objects where object_type = 'PACKAGE'; begin open pkgs; loop fetch pkgs into own, nam; exit when pkgs%notfound; dbms_shared_pool.keep(own || '.' || nam, 'P'); end loop; end; The dbms_shared_pool package was introduced in 7.0 and has evolved over the versions. Until 7.1.5, 'keep' could only be used for packages. Starting with 7.1.6, this was extended to standalone procedures, cursors as well as triggers. For detailed usage instructions, see the dbmspool.sql file. So, prior to this version, if you have large procedures or large anonymous blocks, then these will need to be put into packages and marked kept. With 7.3, most packages do not need to be kept any longer since PL/SQL no longer requires large amounts of contiguous memory to load packages/procedures in memory. You can determine what large stored objects are in the shared pool by selecting from the v$db_object_cache fixed view. This will also tell you which objects have been marked kept. This can be done with the following query: select * from v$db_object_cache where sharable_mem > 10000; Note that this query will not catch PL/SQ: objects that are only rarely used and therefore the PL/SQL object is not currently loaded in the shared pool. To determine what large PL/SQL objects are currently loaded in the shared pool and are not marked 'kept' and therefore may cause a problem, execute the following: select name, sharable_mem from v$db_object_cache where sharable_mem > 10000 and (type = 'PACKAGE' or type = 'PACKAGE BODY' or type = 'FUNCTION' or type = 'PROCEDURE') and kept = 'NO'; Another approach to the above is to use the dbms_shared_pool.sizes procedure. To use this in SQLDBA: set serveroutput on; execute dbms_shared_pool.sizes(10); This should show you the names of all the objects in the shared pool that take more that 10K of memory as well as if they are marked kept or not. For SQL statements, if there are multiple versions of a query (usually a bug if the count is more than 3), they will also be indicated in parenthesis. Use the following query to check for problems: select sql_text, loaded_versions, version_count, sharable_mem from v$sqlarea where loaded_versions > 3 order by sharable_mem; In Oracle7.3 onwards the best candidates for keeping can be seen by querying the table X$KSMSP to see if there are any chunks in the shared-pool that have the KSMCHSIZ larger than 5K and KSMCHCOM like '%PL/SQL%'. If so then one can identify the object name and owner of this chunk using the following SQL: select distinct decode(kglobtyp,0,'CURSOR',7,
'PROCEDURE',8,'FUNCTION',9,'P
ACKAGE', 11,'PACKAGE BODY',12,'TRIGGER',13,'TYPE', 14,'TYPE BODY','OTHER') ||' - '||kglnaown||'.' ||kglnaobj "Eligible PL/SQL objects" from x$kglob where kglobhd4 in (select ksmchpar from x$ksmsp where ksmchcom='PL/SQL MPCODE' and ksmchsiz>5120) If you are 'keeping' PL/SQL objects today and migrate to 7.3 or higher there is no need to re-assess the list of objects that you are keeping. ii) USE BIND VARIABLES One of the best things that can be done to reduce the amount of fragmentation is to reduce or eliminate the number of sql statements in the shared pool that are duplicates of each other except for a constant that is embedded in the statement. The statements should be replaced with one statement that uses a bind variable instead of a constant. For example: select * from emp where empno=1; select * from emp where empno=2; select * from emp where empno=3; Should all be replaced with: select * from emp where empno=:1; You can identify statements that potentially fall into this class with a query like the following: select substr(sql_text, 1, 30) sql, count(*) copies from v$sqlarea group by substr(sql_text, 1, 30) having count(*) > 3; iii) MAX BIND SIZE It is possible for a SQL statement to not be shared because the max bind variable lengths of the bind variables in the statement do not match. This is automatically taken care of for precompiler programs and forms programs, but could be a problem for programs that directly use OCI. The bind call in OCI takes two arguments, one is the max length of the value, and the other is a pointer to the actual length. If the current length is always passed in as the max length instead of the max possible length for the variable, then this could cause the SQL statement not to be shared. To identify statements that might potentially have this problem execute the following statement: select sql_text, version_count from v$sqlarea where version_count > 5; Starting with 7.1.6 this should no longer be an issue as the server can graduate bind buffers even when the user's max bind lengths are jumping up or down and continue to share cursors that are built for larger buffer lengths and flush the smaller sql compilation from the shared pool. iv) ELIMINATING LARGE ANONYMOUS PL/SQL Large anonymous PL/SQL blocks should be turned into small anonymous PL/SQL blocks that call packaged functions. The packages should be 'kept' in memory. For version earlier that 7.3, this includes anonymous PL/SQL blocks that are used for trigger definitions. With 7.3, triggers are compiled and stored to disk like standalone procedures and should be treated as such. Large anonymous blocks can be identified with the following query: select sql_text from v$sqlarea where command_type=47 -- command type for anonymous block and length(sql_text) > 500; Note that this query will not catch PL/SQL blocks that are only rarely used and therefore the PL/SQL block is not currently loaded in the shared pool. Another option that can be used when an anonymous block cannot be turned into a package is to mark the anonymous block with some string so that it can be identified in v$sqlarea and marked 'kept'. For example, instead of using: declare x number; begin x : = 5; end;; you can use: declare /* KEEP_ME */ x number; begin x := 5; end; You can then use the following procedure to select these statements out of the shared pool and mark them
'kept' using the
dbms_shared_pool.keep package. declare /* DONT_KEEP_ME */ addr varchar2(10); hash number; cursor anon is select address, hash_value from v$sqlarea where command_type = 47 -- command type for anonymous block and sql_text like '% KEEP_ME %' and sql_text not like
'%DONT_KEEP_ME%';
begin open anon; loop fetch anon into addr, hash; exit when anon%notfound; dbms_shared_pool.keep(addr ||
',' || to_char(hash), 'C');
end loop; end; v) REDUCING USAGE Another way to reducing fragmentation is to reduce consumption. This is of special importance when using MTS, when every user's session memory is in the shared pool and the impact is multiplied by the total concurrent users. Insert, update, delete and anonymous blocks complete the execution in one round trip. All the memory that is allocated on the server for the execute comes from the PGA and is freed before the call returns to the user. But in the case of selects, memory required to execute the statement - which could be large if a sort was involved - is not freed until the end-of-fetch is reached or the query is cancelled. In these situations using the OCI features to do an exact fetch and cancel helps free memory back to the pool. If the application logic has been embedded into server side PL/SQL, a large number of cursors may be getting cached on the server for every user. Though this results in reduced latch contention and faster response, it does use more memory in the UGA. Setting the close_cached_open_cursors init.ora to TRUE closes the PL/SQL cached cursors on the server, freeing the memory. ***************************** ***************************** *************** 2) COMMON FALLACIES There are a number of common fallacies about the shared pool that are often stated as fact. A) FREE MEMORY One fallacy is that the amount of 'free memory' reported in v$sgastat needs to be kept high. This is incorrect. The free memory reported in this table is not like the free memory reported by operating system statistics. Since the shared pool acts as a cache, nothing will ever be aged out of the shared pool until all the free memory has been used up. This is entirely normal. Free memory is more properly thought of as 'wasted memory'. You would rather see this value be low than very high. In fact, a high value of free memory is sometimes a symptom that a lot of objects have been aged out of the shared pool and therefore the system is experiencing fragmentation problems. B) FLUSH SHARED POOL Some people think that frequently executing 'alter system flush shared_pool' improves the performance of the system and decreases the amount of fragmentation. This is incorrect. Executing this statement causes a big spike in performance and does nothing to improve fragmentation. The only time when it might be useful to run this statement is between shifts of users so that the objects that are relevant to the last shift of users can be flushed out before the next shift of users starts to use the system. This is almost never needed though. ***************************** ***************************** *************** 3) SIZING OF SHARED POOL One very difficult judgement that needs to be make in Oracle7 is to determine the proper size of the shared pool. The following provides some guidelines for this. It should be emphasized that these are just guidelines, there are no hard and fast rules here and experimentation will be needed to determine a good value. The shared pool size is highly application dependent. To determine the shared pool size that will be needed for a production system it is generally necessary to first develop the application and run it on a test system and take some measurements. The test system should be run with a very large value for the shared pool size to make the measurements meaningful. A) OBJECTS STORED IN THE DATABASE The amount of shared pool that needs to be allocated for objects that are stored in the database like packages and views is easy to measure. You can just measure their size directly with the following statement: select sum(sharable_mem) from v$db_object_cache; This is especially effective because all large pl/sql object should be 'kept' in the shared pool at all times. B) SQL The amount of memory needed to store SQL statements in the shared pool is more difficult to measure because of the needs of dynamic SQL. If an application has no dynamic SQL then the amount of memory can simply be measured after the application has run for a while by just selecting it out of the shared pool as follows: select sum(sharable_mem) from v$sqlarea; If the application has a moderate or large amount of dynamic SQL like most applications do, then a certain amount of memory will be needed for the shared SQL, plus more for the dynamic SQL, and more so that the dynamic SQL does not age the shared SQL out of the shared pool. The amount of memory for the shared SQL can be approximated by the following: select sum(sharable_mem) from v$sqlarea where executions > 5; The remaining memory in v$sqlarea is for dynamic SQL Some shared pool will need to be budgeted for this also, but there are few rules here. C) PER-USER PER-CURSOR MEMORY You will need to allow around 250 bytes of memory in the shared pool per concurrent user for each open cursor that the user has whether the cursor is shared or not. During the peak usage time of the production system, you can measure this as follows: select sum(250 * users_opening) from v$sqlarea; In a test system you can measure it by selecting the number of open cursors for a test user and multiplying by the total number of users: select 250 * value bytes_per_user from v$sesstat s, v$statname n where s.statistic# = n.statistic# and n.name = 'opened cursors current' and s.sid = 23; -- replace 23 with session id of user being measured The per-user per-cursor memory is one of the classes of memory that shows up as
'library cache' in v$sgastat.
D) MTS If you are using multi-threaded server, then you will need to allow enough memory for all the shared server users to put their session memory in the shared pool. This can be measured for one user with the following query: select value sess_mem from v$sesstat s, v$statname n where s.statistic# = n.statistic# and n.name = 'session uga memory' and s.sid = 23; -- replace 23 with session id of user being measured A more conservative value to use is the maximum session memory that was ever allocated by the user: select value sess_max_mem from v$sesstat s, v$statname n where s.statistic# = n.statistic# and n.name = 'session uga memory max' and s.sid = 23; -- replace 23 with session id of user being measured To select this value for all the currently logged on users the following query can be used: select sum(value) all_sess_mem from v$sesstat s, v$statname n where s.statistic# = n.statistic# and n.name = 'session uga memory max'; E) OVERHEAD You will need to add a minimum of 30% overhead to the values calculated above to allow for unexpected and unmeasured usage of the shared pool. ***************************** ***************************** *************** 4) FINAL COMMENTS The most important point that needs to be understood by everyone using Oracle7 and PL/SQL (prior to release 7.3) is that all large PL/SQL objects must be made into packages and those packages must be kept in the shared pool. This point cannot be over emphasized. Many customers, especially those running a lot of users, have had terrible performance problems that were completely cleared up by doing this. APPENDIX I: Reserved Shared Pool ============================= ==== 1. RESERVED SPACE FROM THE SHARED POOL ============================= ========= On busy systems, the RDBMS may have difficulty finding a contiguous piece of memory to satisfy a large request for memory. Because the RDBMS will search for and free currently unused memory, the search for this large piece of memory may disrupt the behavior of the share pool, leading to more fragmentation and poor performance. RDBMS 7.1.5 allows DBAs to reserve memory within the shared pool to satisfy these large allocations during RDBMS operations such as PL/SQL compilation and trigger compilation. Smaller objects will not fragment the reserved list, helping to ensure the reserved list will have large contiguous chunks of memory. Once the memory allocated from the reserved list is freed, it returns to the reserved list. The size of the reserved list, as well as the minimum size of the objects that can be allocated from the reserved list are controlled via init.ora parameters: shared_pool_reserved_size and shared_pool_reserved_min_allo c. 1.1 shared_pool_reserved_size ----------------------------- - The init.ora parameter shared_pool_reserved_size controls the amount of shared_pool_size reserved for large allocations. In order to create a reserved list, shared_pool_reserved_size must be greater than shared_pool_reserved_min_allo c. units : bytes default: 0 (no reserved list) minimum: > shared_pool_reserved_min_allo c maximum: 1/2 shared_pool_size 1.2 shared_pool_reserved_min_allo c ----------------------------- ------ The init.ora parameter shared_pool_reserved_min_allo c controls allocation for the reserved memory. Only allocations larger than shared_pool_reserved_min_allo c are allowed to allocate space from the reserved list if a chunk of memory of sufficient size is not found on the shared pool's free lists. units : bytes default: 5000 minimum: 5000 maximum: < shared_pool_reserved_size The default value for shared_pool_reserved_min_allo c should be adequate for almost all systems. 2. CONTROLLING SPACE RECLAMATION OF THE SHARED POOL ============================= ======================= RDBMS 7.1.5 also provides a new procedure, aborted_request_threshold, in package dbms_shared_pool, which allows users to set the limit on the size of allocations allowed to flush the shared pool if the free lists cannot satisfy the request size. Before the RDBMS signals the ORA-04031 error, it incrementally flushes unused objects from the shared pool until there is sufficient memory to satisfy the allocation request. In most cases, incrementally flushing objects from the shared pool frees enough memory for the allocation to complete succesfully. If the RDBMS signals an ORA-04031 error, it has flushed all objects currently not in use on the system without finding a large enough piece of contiguous memory. On a busy system, the larger the space allocation, the more likely the RDBMS will signal the ORA-04031 error. Flushing all objects, however, will impact other users on the system, possibly causing a degradation in performance. The aborted_request_threshold procedure allows the DBA to localize the impact the ORA-04031 error to the process that couldn't allocate memory. The procedure takes a numeric value between 5000 and 2147483647, representing the size, in bytes, of the threshold. 3. NEW FIXED VIEW V$SHARED_POOL_RESERVED ============================= ============ RDBMS 7.1.5 has a new fixed view to help tune the reserved pool and space within the shared pool. The name of the new fixed view is V$SHARED_POOL_RESERVED and has the following columns: Name Null? Type ----------------------------- -- -------- -------------- FREE_SPACE NUMBER AVG_FREE_SIZE NUMBER FREE_COUNT NUMBER MAX_FREE_SIZE NUMBER USED_SPACE NUMBER AVG_USED_SIZE NUMBER USED_COUNT NUMBER MAX_USED_SIZE NUMBER REQUESTS NUMBER REQUEST_MISSES NUMBER LAST_MISS_SIZE NUMBER MAX_MISS_SIZE NUMBER REQUEST_FAILURES NUMBER LAST_FAILURE_SIZE NUMBER ABORTED_REQUEST_THRESHOLD NUMBER ABORTED_REQUESTS NUMBER LAST_ABORTED_SIZE NUMBER These columns of V$SHARED_POOL_RESERVED are only valid if the parameter shared_pool_reserved_size is set to a valid value. FREE_SPACE is the total amount of free space on the reserved list. AVG_FREE_SIZE is the average size of the free memory on the reserved list. FREE_COUNT is the number of free pieces of memory on the reserved list. MAX_FREE_SIZE is the size of the largest free piece of memory on the reserved list. USED_SPACE is the total amount of used memory on the reserved list. AVG_USED_SIZE is the average size of the of the used memory on the reserved list. USED_COUNT is the number of used pieces of memory on the reserved list. MAX_USED_SIZE is the size of the largest used piece of memory on the reserved list. REQUESTS is the number of times that the reserved list was searched for a free piece of memory. REQUEST_MISSES is the number of times the reserved list didn't have a free piece of memory to satisfy the request, and proceeded to start flushing objects from the LRU list. LAST_MISS_SIZE is the request size of the last REQUEST_MISS. MAX_MISS_SIZE is the request size of the largest REQUEST_MISS. The next set of columns contain values which are valid even if shared_pool_reserved_size is not set. REQUEST_FAILURES is the number of times that no memory was found to satisfy a request (example: number of times ORA-04031 occurred) LAST_FAILURE_SIZE is the request size of the last failed request (example: the request size of last ORA-04031). ABORTED_REQUEST_THRESHOLD is the minimum size of a request which will signal an ORA-04031 error without flushing objects. See the procedure aborted_request_threshold described above. LAST_ABORTED_SIZE is the last size of the request which returned an ORA-04031 error without flushing objects from the LRU list. 4. TUNING HINTS BASED ON V$SHARED_POOL_RESERVED ============================= =================== Information in V$SHARED_POOL_RESERVED can help to set values for shared_pool_reserved_size and even shared_pool_size. This section assumes the DBA has performed all other shared pool tuning on his system. 4.1 Initial Value for shared_pool_reserved_size ----------------------------- ------------------- The DBA should make shared_pool_reserved_size 10% of the shared_pool_size. For most systems, this value should be sufficient, if the DBA has already spent time tuning the shared pool. 4.2 Initial Value for shared_pool_reserved_min_allo c ----------------------------- ------------------------ In most cases, the default value for this parameter is adequate. If the DBA increases this value, then the RDBMS will allow fewer allocations from the reserved list and will request more memory from the shared pool list. 4.4 Tuning shared_pool_reserved_size ----------------------------- -------- Ideally, shared_pool_reserved_size should be made large enough to satisfy any request scanning for memory on the reserved list without flushing objects from the shared pool. The amount of operating system memory, however, may constrain the size of the SGA, and therefore the size of the shared pool such that this is not a feasible goal. If the DBA has a system with ample free memory to increase his SGA, the goal is to have: REQUEST_MISS = 0 If the DBA is constrained for OS memory, his goal is: REQUEST_FAILURES = 0 or not increasing LAST_FAILURE_SIZE > shared_pool_reserved_min_allo c AVG_FREE_SIZE > shared_pool_reserved_min_allo c If neither of these goals are met, increase shared_pool_reserved_size; the DBA also needs to increase shared_pool_size by the same amount, since the reserved list is taken from the shared pool. 4.5 shared_pool_reserved_size too low ----------------------------- --------- The reserved pool is too small when: REQUEST_FAILURES > 0 (and increasing) and at least one of the following is true: LAST_FAILURE_SIZE > shared_pool_reserved_min_allo c MAX_FREE_SIZE < shared_pool_reserved_min_allo c FREE_MEMORY < shared_pool_reserved_min_allo c The DBA has two options, depending on his SGA size constraints: o Increase shared_pool_reserved_size and shared_pool_size, accordingly o Increase shared_pool_reserved_min_allo c (but may need to increase shared_pool_size) The first option will increase the amount of memory available on the reserved list without impacting users not allocating memory from the reserved list. The second options reduces the number of allocations allowed to use memory from the reserved list; doing so, however, will increase normal shared pool perhaps impacting other users on the system. 4.6 shared_pool_reserved_size too high ----------------------------- ---------- It is possible that too much memory has been allocated to the reserved list. If: REQUEST_MISS = 0 or not increasing FREE_MEMORY = > 50% of shared_pool_reserved_size minimum The DBA has two options: o Decrease shared_pool_reserved_size o Decrease shared_pool_reserved_min_allo c (if not the default value) 4.7 shared_pool_size too small ----------------------------- -- The new fixed table can also indicate when shared_pool_size is too small. If: REQUEST_FAILURES > 0 and increasing LAST_FAILURE_SIZE < shared_pool_reserved_min_allo c Then the DBA has two options if he has enabled the reserved list: o Decrease shared_pool_reserved_size o Decrease shared_pool_reserved_min_allo c (if set larger than the default) Otherwise, the DBA the could: o Increase shared_pool_size APPENDIX 2: Procedure free_unused_memory ============================= ============ This text is also in the specification for this procedure in dbmsutil.sql. It is part of package dbms_session. Procedure free_unused_memory -- Procedure for users to reclaim unused memory after performing operations requiring large amounts of memory (where large is >100K). Note that this procedure should only be used in cases where memory is at a premium. Examples operations using lots of memory are: o large sorts where entire sort_area_size is used and sort_area_size is hundreds of KB o compiling large PL/SQL packages, procedures, or functions o storing hundreds of KB of data within PL/SQL indexed tables One can monitor user memory by tracking the statistics "session uga memory" and "session pga memory" in the v$sesstat/v$statname fixed views. Monitoring these statistics will also show how much memory this procedure has freed. The behavior of this procedure depends upon the configuration of the server operating on behalf of the client: o dedicated server - returns unused PGA memory to the OS o MTS server - returns unused session memory to the shared_pool In order to free memory using this procedure, the memory must not be in use. Once an operation allocates memory, only the same type of operation can reuse the allocated memory. For example, once memory is allocated for sort, even if the sort is complete and the memory is no longer in use, only another sort can reuse the sort-allocated memory. For both sort and compilation, after the operation is complete, the memory is no longer in use and the user can invoke this procedure to free the unused memory. An indexed table implicitly allocates memory to store values assigned to the indexed table's elements. Thus, the more elements in an indexed table, the more memory the RDBMS allocates to the indexed table. As long as there are elements within the indexed table, the memory associated with an indexed table is in use. The scope of indexed tables determines how long their memory is in use. Indexed tables declared globally are indexed tables declared in packages or package bodies. They allocate memory from session memory. For an indexed table declared globally, the memory will remain in use for the lifetime of a user's login (lifetime of a user's session), and is freed after the user disconnects from ORACLE. Indexed tables declared locally are indexed tables declared within functions, procedures, or anonymous blocks. These indexed tables allocate memory from PGA memory. For an indexed table declared locally, the memory will remain in use for as long as the user is still executing the procedure, function, or anonymous block in which the indexed table is declared. After the procedure, function, or anonymous block is finished executing, the memory is then available for other locally declared indexed tables to use (i.e., the memory is no longer in use). Assigning an uninitialized, "empty," indexed table to an existing index table is a method to explicitly re-initialize the indexed table and the memory associated with the indexed table. After this operation, the memory associated with the indexed table will no longer be in use, making it available to be freed by calling this procedure. This method is particularly useful on indexed tables declared globally which can grow during the lifetime of a user's session, as long as the user no longer needs the contents of the indexed table. The memory rules associated with an indexed table's scope still apply; this method and this procedure, however, allow users to intervene and to explictly free the memory associated with an indexed table. The PL/SQL fragment below illustrates the method and the use of procedure free_unused_user_memory. create package foobar type number_idx_tbl is table of number indexed by binary_integer; store1_table number_idx_tbl; -- PL/SQL indexed table store2_table number_idx_tbl; -- PL/SQL indexed table store3_table number_idx_tbl; -- PL/SQL indexed table ... end; -- end of foobar declare ... empty_table number_idx_tbl; -- uninitialized ("empty") version begin for i in 1..1000000 loop store1_table(i) := i; -- load data end loop; ... store1_table := empty_table; -- "truncate" the indexed table ... -
Received on Tue Jul 25 2000 - 14:21:56 CDT

Original text of this message

HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US