latch: cache buffers chains

原文：https://sites.google.com/site/embtdbo/wait-event-documentation/oracle-latch-cache-buffers-chains

Oracle: latch: cache buffers chains

(article best formated at link:http://tinyurl.com/ybyjazq)   Waits on the cache buffer chains latch, ie the wait event "latch: cache buffers chains" happen when there is extremely high and concurrent access to the same block in a database. Access to a block is normally a fast operation but if concurrent users access a block fast enough, repeatedly then simple access to the block can become an bottleneck. The most common occurance of cbc (cache buffer chains) latch contention happens when multiple users are running nest loop joins on a table and accessing the table driven into via an index. Since the NL  join is basically a   For all rows in i      look up a value in j  where j.field1 = i.val   end loop then table j's index on field1 will get hit for every row returned from i. Now if the lookup on i returns a lot of rows and if multiple users are running this same query then the index root block is going to get hammered on the index j(field1). In order to solve a CBC latch bottleneck we need to know what SQL is causing the bottleneck and what table or index that the SQL statement is using is causing the bottleneck. From ASH data this is fairly easy: select        count(*),        sql_id,        nvl(o.object_name,ash.current_obj#) objn,       substr(o.object_type,0,10) otype,       CURRENT_FILE# fn,       CURRENT_BLOCK# blockn from  v$active_session_history ash     , all_objects o where event like 'latch: cache buffers chains'   and o.object_id (+)= ash.CURRENT_OBJ# group by sql_id, current_obj#, current_file#,                current_block#, o.object_name,o.object_type order by count(*) /                From the out put it looks like we have both the SQL (at least the id, we can get the text with the id) and the block:   CNT SQL_ID        OBJN     OTYPE   FN BLOCKN ---- ------------- -------- ------ --- ------   84 a09r4dwjpv01q MYDUAL   TABLE    1  93170 But the block actually is probably left over from a recent IO and not actually the CBC hot block though it might be. We can investigate further to get more information by looking at P1, P2 and P3 for the CBC latch wait. How can we find out what P1, P2 and P3 mean? by looking them up in V$EVENT_NAME: select * from v$event_name where name = 'latch: cache buffers chains'     EVENT#     NAME                         PARAMETER1 PARAMETER2 PARAMETER3  ---------- ---------------------------- ---------- ---------- ----------         58 latch: cache buffers chains     address     number      tries  So  P1 is the address of the latch for the cbc latch wait.  Now we can group the CBC latch waits by the address and find out what address had the most waits: select     count(*),     lpad(replace(to_char(p1,'XXXXXXXXX'),' ','0'),16,0) laddr from v$active_session_history where event='latch: cache buffers chains' group by p1 order by count(*);    COUNT(*)  LADDR ---------- ----------------       4933 00000004D8108330    In this case, there is only one address that we had waits for, so now we can look up what blocks (headers actually) were at that address select o.name, bh.dbarfil, bh.dbablk, bh.tch from x$bh bh, obj$ o where tch > 5   and hladdr='00000004D8108330'   and o.obj#=bh.obj order by tch NAME        DBARFIL DBABLK  TCH ----------- ------- ------ ---- EMP_CLUSTER       4    394  120         We look for the block with the highest "TCH" or "touch count". Touch count is a count of the times the block has been accesses. The count has some restrictions. The count is only incremented once every 3 seconds, so even if I access the block 1 million times a second, the count will only go up once every 3 seconds. Also, and unfortunately, the count gets zeroed out if the block cycles through the buffer cache, but probably the most unfortunate is that  this analysis only works when the problem is currently happening. Once the problem is over then the blocks will usually get pushed out of the buffer cache. In the case where the CBC latch contention is happening right now we can run all of this analysis in one query select          name, file#, dbablk, obj, tch, hladdr  from x$bh bh     , obj$ o  where         o.obj#(+)=bh.obj and       hladdr in  (     select ltrim(to_char(p1,'XXXXXXXXXX') )     from v$active_session_history      where event like 'latch: cache buffers chains'     group by p1      having count(*) > 5 )    and tch > 5 order by tch    example output NAME          FILE# DBABLK    OBJ TCH HLADDR ------------- ----- ------ ------ --- -------- BBW_INDEX         1 110997  66051  17 6BD91180 IDL_UB1$          1  54837     73  18 6BDB8A80 VIEW$             1   6885     63  20 6BD91180 VIEW$             1   6886     63  24 6BDB8A80 DUAL              1   2082    258  32 6BDB8A80 DUAL              1   2081    258  32 6BD91180 MGMT_EMD_PING     3  26479  50312 272 6BDB8A80                                                  This can be misleading, as TCH gets set to 0 every rap around the LRU and it only gets updated once every 3 seconds, so in this case DUAL was my problem table not MGMT_EMD_PING  Deeper Analysis from Tanel Poder http://blog.tanelpoder.com/2009/08/27/latch-cache-buffers-chains-latch-contention-a-better-way-for-finding-the-hot-block/comment-page-1/#comment-2437 Using Tanel's ideas here's a script to get the objects that we have the most cbc latch waits on col object_name for a35 col cnt for 99999 SELECT   cnt, object_name, object_type,file#, dbablk, obj, tch, hladdr  FROM (   select count(*) cnt, rfile, block from (     SELECT /*+ ORDERED USE_NL(l.x$ksuprlat) */        --l.laddr, u.laddr, u.laddrx, u.laddrr,       dbms_utility.data_block_address_file(to_number(object,'XXXXXXXX')) rfile,       dbms_utility.data_block_address_block(to_number(object,'XXXXXXXX')) block     FROM         (SELECT /*+ NO_MERGE */ 1 FROM DUAL CONNECT BY LEVEL <= 100000) s,        (SELECT ksuprlnm LNAME, ksuprsid sid, ksuprlat laddr,        TO_CHAR(ksulawhy,'XXXXXXXXXXXXXXXX') object         FROM x$ksuprlat) l,        (select  indx, kslednam from x$ksled ) e,        (SELECT                     indx                   , ksusesqh     sqlhash   , ksuseopc   , ksusep1r laddr              FROM x$ksuse) u     WHERE LOWER(l.Lname) LIKE LOWER('%cache buffers chains%')       AND  u.laddr=l.laddr      AND  u.ksuseopc=e.indx      AND  e.kslednam like '%cache buffers chains%'     )    group by rfile, block    ) objs,       x$bh bh,      dba_objects o WHERE        bh.file#=objs.rfile  and  bh.dbablk=objs.block    and  o.object_id=bh.obj order by cnt ; CNT  OBJECT_NAME       TYPE  FILE#  DBABLK    OBJ   TCH  HLADDR ---- ----------------- ----- ----- ------- ------ ----- --------    1 WB_RETROPAY_EARNS TABLE     4   18427  52701  1129 335F7C00    1 WB_RETROPAY_EARNS TABLE     4   18194  52701  1130 335F7C00    3 PS_RETROPAY_RQST  TABLE     4   13253  52689  1143 33656D00    3 PS_RETROPAY_RQST  INDEX     4   13486  52692   997 33656D00    3 WB_JOB            TABLE     4   14443  52698   338 335B9080    5 PS_RETROPAY_RQST  TABLE     4   13020  52689   997 33656D00    5 WB_JOB            TABLE     4   14676  52698   338 335B9080    5 WB_JOB            TABLE     4   13856  52698   338 335F7C00    6 WB_JOB            TABLE     4   13623  52698   338 335F7C00    7 WB_JOB            TABLE     4   14909  52698   338 335B9080  141 WB_JOB            TABLE     4   15142  52698   338 335B9080 2513 WB_JOB            INDEX     4   13719  52699   997 33656D00 Why do we get cache buffers chains latch contention? In order to understand why we get CBC latch contention we have to understand what the CBC latch protects. The CBC latch protects information controlling the buffer cache. Here is a schematic of computer memory and the Oracle processes, SGA and the main components of the SGA: The buffer cache holds in memory versions of datablocks for faster access. Can you imagine though how we find a block we want in the buffer cache? The buffer cache doesn't have a index of blocks it contains and we certainly don't scan the whole cache looking for the block we want (though I have heard that as a concern when people increase the size of there buffer cache). The way we find a block in the buffer cache is by taking the block's address, ie it's file and block number and hashing it. What's hashing? A simple example of hashing is  the "Modulo" function 1 mod 4 = 1 2 mod 4 = 2 3 mod 4 = 3 4 mod 4 = 0 5 mod 4 = 1 6 mod 4 = 2 7 mod 4 = 3 8 mod 4 = 0 Using "mod 4" as a hash funtion creates 4 possible results. These results are used by Oracle as "buckets" or identifiers of locations to store things. The things in this case will be block headers.  Block headers are meta data about data block including pointers to the actual datablock as well as pointers to the other headers in the same bucket.  The block headers in the hash buckets are connected via a doubly linked list. One link points forward the other points backwards The resulting layout looks like the steps to find a block in the cache are If there are a lot of sessions concurrently accessing the same buffer header (or buffer headers in the same bucket) then the latch that protects that bucket will get hot and users will have to wait getting "latch: cache buffers chains" wait. Two ways this can happen (among probably several others) For the nested loops example, Oracle will in some (most?) cases try and pin the root block of the index because Oracle knows we will be using it over and over. When a block is pinned we don't have to use the cbc latch. There seem to be cases (some I think might be bugs) where the root block doesn't get pinned. (I want to look into this more - let me know if you have more info) One thing that can make CBC latch contention worse is if a session is modifying the data block that users are reading because readers will clone a block with uncommitted changes and roll back the changes in the cloned block: all these clone copies will go in the same bucket and be protected by the same latch: How many copies of a block are in the cache? select         count(*)      , name      , file#      , dbablk      , hladdr  from   x$bh bh           , obj$ o where        o.obj#(+)=bh.obj and       hladdr in  (     select ltrim(to_char(p1,'XXXXXXXXXX') )     from v$active_session_history      where event like 'latch: cache%'     group by p1  ) group by name,file#, dbablk, hladdr having count(*) > 1 order by count(*); CNT NAME        FILE#  DBABLK HLADDR --- ---------- ------ ------- --------  14 MYDUAL          1   93170 2C9F4B20 Notice that the number of copies, 14, is higher the the max number of copies allowed set by "_db_block_max_cr_dba = 6" in 10g. The reason is this value is just a directive not a restriction. Oracle tries to limit the  number of copies. Solutions Find SQL ( Why is application hitting the block so hard? ) Possibly change application logic Eliminate hot spots Nested loops, possibly Hash Partition the index with hot block Use Hash Join instead of Nested loop join Use Hash clusters Look up tables (“select language from lang_table where ...”) Change application Use plsql function Spread data out to reduce contention, like set PCTFREE to 0 and recreate the table so that there is only one row per block Select from dual Possibly use x$dual Note starting in 10g Oracle uses the "fast dual" table (ie x$dual) automatically when executing a query on dual as long as the column "dummy" is not accessed. Accessing dummy would be cases like     select count(*) from dual;     select * from dual;     select dummy from dual; an example of not accessing "dummy" would be:     select 1 from dual;     select sysdate from dual; Updates, inserts , select for update on blocks while reading those blocks Cause multiple copies and make things worse What would OEM do? In DB Optimizer: Other References      http://blog.tanelpoder.com/2009/08/27/latch-cache-buffers-chains-latch-contention-a-better-way-for-finding-the-hot-block      http://www.pythian.com/news/1135/tuning-latch-contention-cache-buffers-chain-latches/      http://www.oaktable.net/content/latch-cache-buffer-chains-small-index-primary-key-caused-concurrent-batch-scripts-select-sta#comment-6      http://jonathanlewis.wordpress.com/2008/02/09/index-rebuild-10g/