Professional Documents
Culture Documents
Oracle RAC Performance Tuning
Oracle RAC Performance Tuning
Clusters (http://www.rampant books.com/book_2004_1_10g_grid.htmby Mike Ault and Madhu Tumma. ), The focus of this chapter is on RAC tuning issues. Tuning a RAC database is very similar to tuning a non -RAC database; however, t here are some major differences that will be covered. This chapter will also cov er the views used to monitor the RAC environment. Finally, the Oracle Enterprise Manager features that automate some of the DBAs RAC monitoring tasks will be int roduced. Analysis of Per formance Issues The analysis of performance issues in RAC involves several key areas: n n n n Normal database tuning and monitoring. Monitoring RAC cluster interconne ct performance. Monitoring workloads. Monitoring RAC - specific contention. Normal database monitor ing is covered thoroughly in any number of other texts. Thus, aspects of database tuning such as SQL tuning or standard SGA and internal s tuning are not covered in this text other than the required extensions to norm al database monitoring. Monitoring RAC C luster Interconnect Performance The most important aspects of RAC tuning are the monitoring and tuning of the gl obal services directory processes. The processes in the (GSD) communicate throug h the cluste Global Service Daemon r interconnects. If the cluster interconnects do not perform properly, the entire RAC structure will suffer no matter how well everything else is tuned. The major processes of concern are the Global Enqueue Services (GES) an d Global Cache Services (GCS) processes.
v$session_wait, v$obj_stats, and v$enqueues_stats. The major waits to be concern ed with for RAC are: n n n global cache busy buffer busy global cache buffer busy global cr In later versions of Oracle, the global cache is shortened to just gc. To find t he values for these waits, the gv$session_wait view is used to identify objects that have performance issues. The gv$session_wait view contents are shown in the following results: Description of GV$SESSION_WAIT Name Null? ---------------------------------------- -------INST_ID SID SEQ# EVENT P1TEXT P1 P1RAW P2TEXT P2 P2RAW P3TEXT P3 P3RA W WAIT_CLASS# WAIT_CLASS WAIT_TIME SECONDS_IN_WAIT STATE Type ------------NUMBER NUMBER NUMBER VARCHAR2(64) VARCHAR2(64) NUMBER RAW(4) VARCHAR2(64) NUMBER RAW(4 ) VARCHAR2(64) NUMBER RAW(4) NUMBER VARCHAR2(64) NUMBER NUMBER VARCHAR2(19) wait_class column which is used to restrict returned values based on 12 basic wa it classes, one of which is the cluster wait class. New in 10g is the The follo wing wait events indicate that the remotely cached blocks were shipped to the local instance without having been busy, pinned or requiring a log flush and can safely be ignored: n n n n gc current block 2-way gc current block 3-way gc cr block 2-way gc cr block 3-wa y However, the object level statistics for gc current blocks received and gc cr bl ocks received enable the rapid identification of the indexes and tables which ar e shared by the active instances.
p1 and p2 identify the file and block number of any object experiencing the abov e waits for the events, as shown in the f ollowing queries: SELECT INST_ID, EVENT, P1 FILE_NUMBER, P2 BLOCK_NUMBER, WAIT_TIME FROM GV$SESSIO N_WAIT WHERE EVENT IN ( buffer busy global cr , global cache busy , buffer bus y global cache ); The output from this quer y should resemble the following: INST_ID ------1 2 EVE NT FILE_NUMBER BLOCK_NUMBER WAIT_TIME ----------------------- ----------- ----------- ---------global cache busy 9 150 0 global cache busy 9 150 0 In order to find the object that corresponds to a particular file and block, the following query can be issued for the first combination on the above list: SELE CT OWNER, SEGMENT_NAME, SEGMENT_TYPE FROM DBA_EXTENTS WHERE FILE_ID = 9 AND 150 BETWEEN BLOCK_ID AND BLOCK_ID+BLOCKS-1; SEE CODE DEPOT FOR COMPLETE SCRIPT (http ://www.dbaoracle.com/bp/bp_elec_adv_mon_tuning.htm) In this example, there is no need to worry about the instance as this SELECT is issued from within the cluster because they all see the same tables, indexes and objects. The output will be similar to: OWNER SEGMENT_NAME SEGMENT_TYPE ---------- --------------------------- --------------SYSTEM MOD_TEST_IND INDEX Once the objects causing the contention are determined, they should be modified by: n n n Reducing the rows per block. Adjusting the block size. Modifying initrans and freelists.
Index leaf blocks are usually the most contended objects in the RAC environment; therefore, using a smaller block size for index objects can decrease intra -ins tance contention for index leaf blocks. Contention in blocks can be measured by using the block transfer time. To determine block transfer time, examine the sta tistics g lobal cache cr block receive time and global cache cr blocks received. The time is determined by calculating the ratio of global cache cr block receiv e time to global cache cr blocks received. The values for these statistics are t aken from the gv$sysstat view shown below: Description of GV$SYSSTAT Name Null? ----------------------------------------- ------INST_ID STATISTIC# NAME CLASS VALUE HASH Type ------------NUMBER NUMBER VA RCHAR2(64) NUMBER NUMBER NUMBER The following script shows this calculation. column "AVG RECEIVE TIME (ms)" form at 9999999.9 col inst_id for 9999 prompt GCS CR BLOCKS select b1.inst_id, b2.val ue "RECEIVED", b1.value "RECEIVE TIME", ((b1.value / b2.value) * 10) "AVG RECEIV E TIME (ms)" from gv$sysstat b1, gv$sysstat b2 where SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.dbaoracle.com/bp/bp_elec_adv_mon_tuning.htm) INST_ID RECEIVE D RECEIVE TIME AVG RECEIVE TIME (ms) ------- ---------- ------------ -------------------1 2791 3287 14.8 2 3760 7482 19.9 If the transfer time is too high, or if one of the nodes in the cluster shows ex cessive transfer times, the cluster interconnects should be checked using system level commands to verify that they are functioning correctly. In the above SELECT result, instance two exhibits an average receive time that is 69% higher than instance one. SELECT for determini ng latency measures the overall latency, including that for queue, build, flush and send time. These statistics are also found in the gv$sy sstat view. The SELECT is shown below: Another useful SELECT a.inst_id "Instance", (a.value+b.value+c.value)/d.value "L MS Service Time" FROM GV$SYSSTAT A,
GV$SYSSTAT B, GV$SYSSTAT C, GV$SYSSTAT D WHERE A.name = global cache cr block b uild time AND B.name = global cache cr block flush time AND C.name = global cache cr block send time AND D.name = global cache cr blocks served AND SEE C ODE DEPOT FOR COMPLETE SCRIPT (http://www.dbaoracle.com/bp/bp_elec_adv_mon_tunin g.htm) Instance LMS Service Time --------- ---------------1 1.07933923 2 .636687 318 Why does instance two show a lower service time than instance one? The node cont aining instance two is the slower node, while the node containing instance one s erves blocks faster. The reason is that the slower node two serves blocks to nod e one, so node one shows a slower service time even though it is a faster node! Be aware of this in a two node cluster. Where there are multiple nodes, the serv ice times should average out. If not, the aberrant node is probably troubled. The following code can be used to examine the individual components of the servi ce time to determine the source of the problem: SELECT A.inst_id "Instance", (A. value/D.value) "Consistent Read Build", (B.value/D.value) "Log Flush Wait", (C.v alue/D.value) "Send Time" FROM GV$SYSSTAT A, GV$SYSSTAT B, GV$SYSSTAT C, GV$SYSS TAT D WHERE A.name = global cache cr block build time AND B.name = global cac he cr block flush time AND C.name = global cache cr block send time AND D.nam e = global cache cr blocks served AND SEE CODE DEPOT FOR COMPLETE SCRIPT (http ://www.dbaoracle.com/bp/bp_elec_adv_mon_tuning.htm) Instance Consistent Read Bui ld Log Flush Wait Send Time --------- --------------------- -------------- --------1 .00737234 1.05059755 .02203942 2 .04645529 .51214820 .07844674 If problems are detected, operating system specific commands should be used to p inpoint the node having difficulties. For example, in the SUN Sol aris environment, commands such as the following can be used: netstat -l netstat -s s ar -c sar -q vmstat
n A large number of processes in the run queue waiting for CPU or scheduling delay s. n Platform specific operating system parameter settings that affect IPC buffering or process scheduling. n Slow, busy or faulty interconnects. In these cases, look for dropped packets, re transmits, or cyclic redundancy check (CRC) errors. Ensure that the netwo is pri vate and that inter - instance traffic is not routed through a public network. r k The DBA will be interested in the waits and statistics for tuning shown in Table 14.1. Statistic or Wait buffer busy global cache buffer busy waits cache convert waits cache open waits consistent gets cr request retry db block changes db block gets Description A wait event that is signaled when a process waits for a block to be available because another process is already obtaining a resource for this bloc k. A wait event that is signaled when a process cannot get a buffer due to anoth er process using the buffer at that point in time. A statistic showing the total number of waits for all up-convert operations, these are: global cache null to S, global cache null to X, and global cache S to X. A statistic that shows the t otal number of waits for global cache open S and global cache open X. The consis tent gets statistic shows the number of buffers obtained in consistent read (CR) mode. A statistic that quantifies when Oracle resubmits a consistent read reque st due to detecting that the holding instance is no longer available. The number of current buffers that where obtained in exclusive mode for DML. The number of current buffers that were obtained for a read. Source gv$system_event, gv$session_wait gv$system_event, gv$session_wait gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat
Statistic or Wait DBWR crossinstance writes Description These are also forced writes and show the number of writes that an i nstance had to perform to disk in order to make a previously exclusively held bl ock available for another instance. Generally, DBWR cross-instance writes are el iminated due to Cache Fusion unless a value greater than 0 (zero) is set for the gc_files_to_locks parameter. global cache bg This wait event can only occur dur ing acks the startup or shutdown of an instance as the LMS process completes its operations. global cache This is a wait event that accumulates busy when a sess ion has waits for an ongoing operation on the resource to complete. global cache cr This is a wait event that accumulates cancel wait when a session waits for t he completion of an AST for a canceled block access request. The canceling of th e request is part of the Cache Fusion Write Protocol. global cache This is a sta tistic that shows the converts resource converts for buffer cache blocks. Whenev er GCS resources are converted from NULL to SHARED, NULL to EXCLUSIVE or SHARED to EXCLUSIVE this statistic is incremented. global cache This statistics shows t he accumulated convert time time for global conversions on GCS resources for all sessions. global cache Whenever a resource operation times convert timeouts out this statistic is incremented. global cache cr This statistic shows the time wa ited for block flush time a log flush whenever a CR request is served. This is p art of the serve time. global cache cr If a process requests a consistent read b locks received for a data block that is not in its local cache it has to send a request to another instance, this statistic is incremented once the buffer has b een received. global cache cr This statistic records the accumulated block recei ve round trip time for all requests for time consistent read blocks. Source gv$sysstat gv$system_event, gv$session_wait gv$system_event, gv$session_wait gv$system_even t, gv$session_wait gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat
Statistic or Wait global cache cr blocks served Description This statistic shows the number of requests for a consistent read bl ock served by LMS. This statistic is incremented by Oracle when the block is sen t. global cache cr This statistic shows the accumulated block build time time th at the LMS process needs to create a consistent read block on the holding instan ce global cache cr This statistic shows the time needed block send time by LMS t o begin a send of a consistent read block. For send time, timing does not stop u ntil send has completed. Only the time it takes to initiate the send is measured ; the time elapsed before the block arrives at the requesting instance is not in cluded. global cache cr This is a wait event that happens when cancel wait a ses sion waits for a canceled CR request to complete its acquisition interrupt. The process of canceling the CR request is part of the Cache Fusion Write Protocol. global cache cr This is a wait event that happens request because a process is m ade to wait for a pending CR request to complete. The process must wait for eith er shared access to a block to be granted before it can read the block from disk , or it waits for the LMS of the holding instance to send the block. global cach e This is a statistic that shows the time it current block takes to flush the ch anges from a block flush time to disk before the block can be shipped to the req uesting instance. global cache This is a statistic that shows the time it curren t block pin takes to pin a current block before time shipping it to a requesting instance. The pinning of a block prevents further changes to the block while it is prepared to be shipped. global cache This is a statistic that shows the curr ent blocks number of current blocks received over received the interconnect from the holding instance. global cache This is a statistic that shows the current b lock current blocks accumulated round trip receive time time for all requests. Source gv$sysstat gv$sysstat gv$sysstat gv$system_event, gv$session_wait gv$system_event, gv$session_wait gv$sysstat gv$sysstat gv$sysstat gv$sysstat
Statistic or Wait global cache current block send time Description This is a statistic that shows the time it takes to send the current block over the interconnect to the requesting instance. global cache This is a statistic that shows the current blocks number of current blocks shipped over se rved the interconnect to the requesting instance. global cache This is a statist ic that shows Oracle freelist wait must wait after it detects that the local ele ment free list is empty. global cache This is a statistic that shows the freelis t waits number of times that the resource element free list was found empty by O racle. global cache This is a statistic that shows the gets number of buffer get s that caused the opening of a new GCS resource. global cache get This is a stat istic that shows the time accumulated time for all sessions that was needed to o pen a GCS resource for a local buffer. global cache A wait event that occurs whe never a null to S session has to wait for a resource conversion to complete. glo bal cache This is a wait event that occurs null to X whenever a session has to w ait for this type of resource conversion to complete. global cache This is a wai t event that occurs when a open S session has to wait for receiving permission f or shared access to the requested resource. global cache This is a wait event th at occurs when a open X session has to wait for receiving exclusive access to th e requested resource. global cache S This is a wait event that occurs to X whene ver a session has to wait for this type of resource conversion to complete. glob al cache This is a wait event that can occur pending ast when a process waits fo r an acquisition interrupt before Oracle closes a resource element. Source gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$system_event, gv$session_wait gv$system_event, gv$session_wait gv$system_even t, gv$session_wait gv$system_event, gv$session_wait gv$system_event, gv$session_ wait gv$system_event, gv$session_wait
Statistic or Wait Description global cache This is a wait event that occurs when a pred cancel wait session must wait for the acquisition interrupt to complete for a canceled predecessor read request. The canceling of a predecessor read req uest is a part of the Cache Fusion Write Protocol. global cache This is a wait e vent that occurs retry prepare whenever Oracle cannot ignore or skip a failure t o prepare a buffer for a consistent read or Cache Fusion request. global lock Th is is a statistic showing the number async converts of resource converts from an incompatible mode. global lock sync This is a statistic showing the number gets of synchronous GCS resources that Oracle must open. These sync gets are usually for GES resources. Library cache resources are one example. global lock This is a statistic showing the number async gets of Ansynchronous GES resources that O racle must open. Normally, async gets include the number of global cache gets an d are only used for GES resources. global lock get This is a statistic that show s the time accumulated open time for all GES resources. global lock sync This st atistic shows the number of converts GES resources that Oracle converted from an incompatible mode. Usually sync converts occur for GES resources only. global l ock This is a statistic that shows the convert time accumulated global lock sync converts and global lock async converts time. lock buffers for This is a statis tic that shows the read number of NULL to SHARED up converts. lock gets per This is a statistic that shows the transaction number of global lock sync gets and g lobal lock async gets per transaction. lock converts This is a statistic that sh ows the per transaction number of global local sync converts and global lock asy nc converts per transaction. Source gv$system_event, gv$session_wait gv$system_event, gv$session_wait gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat gv$sysstat
Statistic or Wait messages flow controlled messages received messages sent direc tly messages sent indirectly physical reads physical writes remote instance undo block writes remote instance undo header writes Table 14.1: Description This shows the number of messages that are intended to be sent direc tly but that were instead queued and delivered later by LMD/LMS. This statistics shows the number of messages received by the LMD process. This statistic shows the number of messages sent directly by Oracle processes. This statistics shows the number of explicitly queued messages. This is a statistic that shows the tot al number of disk reads performed because a request for a data block could not b e satisfied from a local cache. This statistics shows the number of write I/Os p erformed by the DBWNn processes. This number includes the number of DBWR cross i nstance writes (forced writes) in Oracle Database 10g when gc_files_to_locks is set. Setting gc_files_to_locks for a particular datafile will enable the use of the old ping protocol and will not leverage the Cache Fusion architecture. This is a statistic that shows the number of undo blocks written to disk by DBWn due to a forced write. This is a statistic that shows the number of rollback segment header blocks written to disk by DBWn due to a forced write. Source gv$dml_misc gv$dml_misc gv$dml_misc gv$dml_misc gv$sysstat gv$sysstat gv$sysstat gv$sysstat Waits and Statistics for RAC Monitoring from Oracle RAC Manual Glossary Oracle has provided the racdiag.sql script that can help troubleshoot RAC slowdo wns; however, it has limited applicability to RAC system tuning. It is suggested that STATSPACK be used, combined with the selects provided in this chapter for RAC tuni ng. racdiag.sql is available at metalink.oracle.com . The class column in the gv$sysstat view indicates the type of statistic. Oracle RAC related statistics are in classes eight, 32, and 40.
n global cache blocks lost: This statistic shows block losses during transfers. Hi gh values indicate network pr oblems. The use of an unreliable IPC protocol such as UDP may result in the value for global cache blocks lost being non -zero. Wh en this occurs, take the ratio of global cache blocks lost divided by global cac he current blocks served plus global cache cr blocks served. This ratio should b e as small as possible. Many times, a non -zero value for global cache blocks lo st does not indicate a problem because Oracle will retry the block transfer oper ation until it is suc cessful. global cache blocks corrupt: This statistic shows if any blocks were corrupted during transfers. If high values are returned for this statistic, there is probably an IPC, network, or hardware problem. n An exampl e SELECT to determine if further examination is needed would be: SELECT A.VALUE "GC BLOCKS LOST 1", B.VALUE "GC BLOCKS CORRUPT 1", C.VALUE "GC BL OCKS LOST 2", D.VALUE "GC BLOCKS CORRUPT 2" FROM GV$SYSSTAT A, GV$SYSSTAT B, GV$ SYSSTAT C, GV$SYSSTAT D WHERE A.INST_ID=1 AND A.NAME= gc blocks lost AND B.INST _ID=1 AND B.NAME= gc blocks corrupt AND C.INST_ID=2 AND C.NAME= gc blocks lost AND D.INST_ID=2 AND D.NAME= gc blocks corrupt ; A sample result from the above select should l ook like the following: GC BLOCKS LOST 1 GC BLOCKS CORRUPT 1 GC BLOCKS LOST 2 GC BLOCKS CORRUPT 2 --------------- ------------------- ---------------- ------------------0 0 652 0 In this result, instance 2 is showing some problems with lost blocks, so it migh t be useful to look at the ratio described above: SELECT A.INST_ID "INSTANCE", A .VALUE "GC BLOCKS LOST", B.VALUE "GC CUR BLOCKS SERVED", C.VALUE "GC CR BLOCKS S ERVED", A.VALUE/(B.VALUE+C.VALUE) RATIO FROM GV$SYSSTAT A, GV$SYSSTAT B, GV$SYSS TAT C WHERE A.NAME= gc blocks lost AND B.NAME= gc current blocks served AND C. NAME= gc cr blocks served and SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.db aoracle.com/bp/bp_elec_adv_mon_tuning.htm) Instance gc blocks lost gc cur blocks served gc cr blocks served RATIO --------- -------------- -------------------- ------------------- ---------1 0 3923 2734 0 2 652 3008 4380 .088251218
db_file_multiblock_read_count of 16, this was causing excessive network traffic because the system was using full table scans resulting in a read of 128K. In ad dition, the actua l TCP buffer area was set to a small number. Setting these val ues for the TCP receive and send buffers is an operating specific operation. DBAs should talk this over with the system operator or check out: http://www.psc.edu/networking/perf_tune.html The following section covers example commands for various operating systems. DEC Alpha (Digital UNIX) $ dbx -k /vmunix (dbx) assign sb_max = (u_long) 524288 (dbx) patch sb_max = (u_long) 524288 T he first command changes the value for the operating system and the second pat ches the value into the kernel so it will not be lost at the next reboot. mbclus ters can be modified to at least 832, and the kernel variables tcp_sendspace and tcp_ recvspace cam also be modified in this manner. Under version 4.0, use the syscon fig -r inet <variable> <value> command to do this type of modification. HPUX (11 ) Use the ndd command to view and modify the various TCP variable h <parm_name> to get the method for changing the values. AIX Use the "no" command. Linux To ch eck the values you CAT the contents of the various $cat /proc/sys/net/core/rmem_default s. Use the ndd /proc/sys/net files: The values of interest are: rmem_default rmem_max
wmem_default wmem_max tcp_rmem tcp_wmem The value under Linux can be set by echoing the proper values into the files in /proc/sys/net. Some example values that have been found to be beneficial are: echo 100 5000 640 2560 150 30000 5000 1884 2 >/proc/sys/vm/bdflush hdparm -m16 -c1 -d1 -a8 /dev/hda hdparm -m16 -c1 -d1 -a8 /dev/hdb echo 131071 >/proc/sys/ne t/core/rmem_default echo 262143 >/proc/sys/net/core/rmem_max echo 131071 >/pro c/sys/net/core/wmem_default echo 262143 >/proc/sys/net/core/wmem_max echo 4096 65536 4194304 >/proc/sys/net/ipv4/tcp_wmem echo 4096 87380 4194304 >/proc/sys/ net/ipv4/tcp_rmem /proc/sys/vm/bdflush actually improves memory handling. The two hdparm commands improve IDE disk performances. The most i nteresting ones here are the ones echo ing values into /proc/sys/net/core and /proc/sys/net/ipv4 subdirectory files. The call to SUN (2.x) SUN uses the ndd command: $ndd -set /dev/tcp tcp_max_buff xxx $ndd -set /dev/tcp tcp_xmit_hiwat xxx $ndd set /dev/tcp tcp_recv_hiwat xxx Windows NT/2000 Use the various settings in the registry under the key: HKEY_LOCAL_MACHINE \SYSTEM\CurrentControlSet\Services\Tcpip\Parameters Additional detailed notes are available at: http://rdweb.cns.vt.edu/public/notes/win2k-tcpip.htm After increasing both the buffer size and buffer space for the TCP area, the fol lowing results were obtained with the query above after the same series of tests were repeated: INSTANCE G C BLOCKS LOST GC CUR BLOCKS SERVED GC CR BLOCKS SERVED RATIO --------- -------------- -------------------- ------------------- ---------1 0 3892 364 2 0 2 0 3048 2627 0 By correcting the TCP issue, performance increased such that the tests now
required only three minutes 20 seconds to complete on the poor performing node.
current Blocks In addition to blocks, RAC performance is also a concern when pro cessing cr current mode blocks. Current mode blocks suffer from latency as well as build an d wait time concerns similar to cr blocks. The average latency for a current mod e block is calculated with the SELECT: column "AVG RECEIVE TIME (ms)" format 9999999.9 col inst_id for 9999 prompt GCS CURRENT BLOCKS select b1.inst_id, b2.value "RECEIVED", b1.value "RECEIVE TIME", ((b1.value / b2.value) * 10) "AVG RECEIVE TIME (ms)" from gv$sysstat b1, gv$syss tat b2 where b1.name = gc current block receive time and b2.name = gc current blocks received and b1.inst_id = b2.inst_id; INST_ID RECEIVED RECEIVE TIME AVG RECEIVE TIME (ms) ------ ---------- ------------ --------------------1 22694 68 999 30.4 2 23931 42090 17.6 The service time for receiving a current block is calculated in a similar fashio n to the value for acr block, except there is a pin time instead of a build time: SELECT a.inst_id "Instance", (a.value+b.value+c.value)/d.value "Current Blk Serv ice Time" FROM GV$SYSSTAT A, GV$SYSSTAT B, GV$SYSSTAT C, GV$SYSSTAT D WHERE A.na me = gc current block pin time AND B.name = gc current block flush time AND C.name = gc current block send time AND D.name = gc current blocks served AN D SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.dbaoracle.com/bp/bp_elec_adv_mo n_tuning.htm) Instance Current Blk Service Time --------- -----------------------1 1.18461603 2 1.63126376 Instance two is requiring more time to service current blocks. of the problem de termined? The overall service time can be How is the source decomposed t o determine where the area of concern lies: SELECT A.inst_id "Instance", (A.value/ D.value) "Current Block Pin", (B.value/D.value) "Log Flush Wait", (C.value/D.val ue) "Send Time" FROM
GV$SYSSTAT A, GV$SYSSTAT B, GV$SYSSTAT C, GV$SYSSTAT D WHERE A.name = gc curren t block build time AND B.name = gc current block flush time AND C.name = gc current block send time AND D.name = gc current blocks served AND SEE CODE DE POT FOR COMPLETE SCRIPT (http://www.dbaoracle.com/bp/bp_elec_adv_mon_tuning.htm) Instance Current Block Pin Log Flush Wait Send Time --------- ------------------------------ ---------1 .69366887 .472058762 .018196236 2 1.07740715 .4805491 99 .072346418 In this case, most of the time difference comes from the pin time for the curren t block in instance two. High pin times coul d indicate problems at the I/O inte rface level. A final set of statistics deals with the average global cache convert time and the average global cache get times. The following SELECT can be used to get this information from the RAC database: select a.inst_id "Instance", a.value/b.value "Avg Cache Conv. Time", c.value/d.v alue "Avg Cache Get Time", e.value "GC Convert Timeouts" from GV$SYSSTAT A, GV$S YSSTAT B, GV$SYSSTAT C, GV$SYSSTAT D, GV$SYSSTAT E where a.name= gc convert time and b.name= gc converts and c.name= gc get time and d.name= gc gets and e.n ame= gc convert timeouts and b.inst_id=a.inst_id and c.inst_id=a.inst_id and d. inst_id=a.inst_id and e.inst_id=a.inst_id order by a.inst_id; SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.dbaoracle.com/bp/bp_elec_adv_mon_tuning.htm) Instance Avg Cache Conv. Time Avg Cache Get Time GC Convert Timeouts --------- ------------------- ------------------ ------------------1 1.85812072 .981296356 0 2 1.65947528 .627444273 0 For th is database, instance one has the highest convert and get times, as expec ted, since it is converting and getting from instance two, which is the slow ins tance. None of the times are excessive, >10 -20 ms.
n Hi gh convert times indicate excessive global concurrency requirements. In other words, the instances are swapping a lot of blocks over the interconnect. n Large values or rapid increases in the gets, converts, or average times indicate GCS contention. n n Latenci The es for resource operations may be high due to excessive system loads. gv$system_event view can be used to review the time_waited statistics for variou s GCS events if the get or convert times become significan t. STATSPACK is good for this. Values other than zero for the GC converts timeouts indicate system contention o r congestion. Timeouts should not occur and indicate a serious performance probl em. n Additional Wait Events of Concern Most of the events report ed in the dynamic performance views or in a STATSPACK report that show a high total time are actually normal. However, if mo nitored response times increase and the STATSPACK report shows a high proportion
of wait time for cluster accesses, the cause of the determined. STATSPACK repor ts provide a breakdown of the wait events, with the five highest values sorted b y percentages. The following specific RAC events should be monitored: n n n n n - related se waits needs to be global cache open s: A block was selected global cache open x: A block was selec ted . for IUD. SELECT. global cache null to s: A block was transferred for global cache null to x: A blo ck was transferred for IUD. global cache cr reques t: A block was requested transferred for consistent read purposes. Global Cache Service Utilization for Logical Reads. n these events to help n The following sections will provide more information o show why they are importa nt to monitor. The global cache open s and global cache open x Events The initial access of a particular data block by an instance generates these eve nts. The duration of the wait should be short, and thepletion of the wait is mos t com likely followed by a read from disk. This wait is a result of the blocks t hat are being requested and not being cached in any instance in the cluster data base. This necessitates a disk read.
buffer cache hit ratios may also be observed. Unfortunately, other than preloadi ng heavily used tables into the buffer caches, there is little that can be done about this type of wait event. The global cache null to s and global cache null to x Events These events are generated by inter -instance block ping across the network. -In ter instance block ping is when two instances exchange the same block back and f orth. Processes waiting for cache null to s events are waiting for a block to be global transferred from the instance that last changed it. When one instance repeatedly requests cached data blocks from the other RAC instances, these events consume a greater proportion of the total wait time. The only method for reducing these events is to reduce the number of rows per block to eliminate the need for block swapping between two instances in the RAC cluster. The global cache cr request Event This event is generated when an instance has requested a consistent read da ta b lock and the block to be transferred had not arrived at the requesting instance. Other than examining the cluster interconnects for possible problems, there is nothing that can be done about this event other than to modify objects to reduce the possibi of contention. lity Global Cache Service Times When global cache wa its constitute a large proportion of the wait time, as listed on the first page of the STATSPACK or AWRRPT report, and if response time or throughput does not c onform to service level rements, the Global Cache requi Service workload charact eristics on the cluster statistics page of the STATSPACK or AWRRPT reports shoul d be examinedThe STATSPACK or AWRRPT . reports should be taken during heavy RAC workloads. If the STATSPACK report shows that average GCS time per request is hi gh, it the is the result of one of the following: n n n Contention for blocks. System loads. Network issues. The operating system logs and operating system statistics are used to indicate w hether a network link is congested. A network link can be congested if:
n Packets are being routed through the public network instead of the private inter connect. n The sizes of the run queues are increasing. If CPU usage is maxed out and processes are queuing for the CPU, the priority of the GC S processes (LMSn ) can be raised over other processes to lower GCS times. The load on the server can also be alleviated by reducing the number of p rocesses on the database server, increasing capacity by adding CPUs to the serve r, or adding nod es to the cluster database. New 10g RAC Waits This section covers the most important Oracle Database 10g wait events to be awa re of when analyzing performance reports. When user response times increase and a high proportion of time waited is due to the global cache (gc), the cause must be determined and corrected. It is best to start with an ADDM report if that opt ion has been purchased. This will analyze the routinely collected performance st atistics with respect to their impact pinpoint the objects and SQL contributing most to the time waited. After that, the analysis can move on to the more detail ed reports produced by the AWR and STATSPACK utilities. The most important wait events for RAC fall into four main categories, arranged alphabetically these are : n Block n n n n n - oriented waits: gc current block 2-way gc current block 3-way gc cr block 2-way gc cr block 3-wa y - oriented waits: Contention n n n gc current block busy gc cr block busy gc current buffer busy gc current block c ongested gc cr block congested n Load-oriented waits: n n n Message-oriented waits:
n n gc current grant 2-way gc cr grant 2-way Block Oriented Wait Events As their names imply, these block related wait events are the result of some combinati on of message and transfer event. Whether the wait occurred between a resource master and a requesting node as a single messag e and block transfer or was the result of a message passed from a node several t imes removed from the resource master causing two messages and a block transfer, the result is the same. A wait event of some duration occurred. These wait even ts consist of several components: n n n Physical network latency Request processing time Requesting process sleep time By looking at the total and ave rage wait times associated with these events, the DBA can be alerted when performance is being affected in a negative manner, such as when average times increase beyond statistical norms, something is wrong and needs investigation. Usually, the impact of i issues or the interactions of nod es against hot blocks are the culprit when these waits become significant. nterc onnect malfunctions or load Message Related Wait Events The wait events dealing with messages usually indicate that the block requested was not present in any of the caches for the various nodes. This results in the need to issue a global grant and allow the requesting instance to read the block from disk in order to modify it. When messaging events result in high waits, it is usually due to fre too small, in the event of quently accessed SQL causing d isk IO, due to cache areas being cr grants, or the instances are inserting a lot of new records requiring format of disk blocks, if current grants are the cause. Contention -Oriented Wait Event s Conte ntion -oriented wait events are caused by: n n n n Waits for blocks pinned by other nodes Waits for changes to be flushed t o disk Waits due to high concurrency for block access Intra - node cache - fusion conte ntion
n n n n Increasing the available CPUs Load balancing Offloading processing until a lower load is available Adding a n ew node - trip, beginning For load - oriented events, the time is the total time for the round when the request is made and ending when the block is received. The following section will cover the STATSPACK report sections that are useful w hen dealing with RAC . The report is only in relation to the node/instance on statspack.snap procedure and spreport.sql report script can be run on each node/ instance that is to be m onitored. which it was run. To compare other instances, the RAC and STATSPACK The following is a STATSPACK report for the troubled instance. The sections covered will be those that involve RAC statistics. The first secti on deals with the top five timed events: Top 5 Timed Events ~~~~~~~~~~~~~~~~~~ % Total Event Waits Time (s) Ela Time ------------------------------------------- ------------ ----------- -------global cache cr request 820 154 72.50 CPU time 54 25.34 global cache null to x 478 1 .5 2 control file sequential read 600 1 .52 control file parallel write 141 1 .28 -----------------------------------------------------------% total elapsed time column that are greater than or near the %total ela time va lue for cpu time. The cpu time statistic should be the predominant event as it d enotes processing time. If cpu time is not the predominant event, the events tha t exceed cpu times share of the tota l elapsed time need to be investigated. In t he above report section, global cache cr request events are dominating the repor t. This indicates that transfer times are excessive from the other instances in the cluster to this instance. The excessive transfer times could be due to netwo rk problems or buffer cache sizing issues. Observe the events in the report that are taking a majority of the
Top 5 Timed Events ~~~~~~~~~~~~~~~~~~ % Total Event Waits Time (s) Ela Time ------------------------------------------- ------------ ----------- -------CPU tim e 99 64.87 global cache null to x 1,655 28 18.43 enqueue 46 8 5.12 global cache busy 104 7 4.73 DFS lock handle 38 2 1.64 cpu time, followed by global cache null to x, which indicates the major wait has been shifted from in tra-cache to I/O -based as global cache null to x indicate s a read from disk. The number one wait is now The next report in the STATSPACK listing shows the workload characteristics for the instance for which the report was generated: Cluster Statistics for DB: MIKE Instance: mike2 Snaps: 25 -26 Global Cache Service - Workload Characteristics ---------------------------------------------Ave global cache get time (ms): Ave global cache convert time (ms) : Ave Ave Ave Ave Ave Ave Ave Ave Ave Ave build time for CR block (ms): flush ti me for CR block (ms): send time for CR block (ms): time to process CR block requ est (ms): receive time for CR block (ms): pin time for current block (ms): flush time for current block (ms): send time for current block (ms): time to process current block request (ms): receive time for current block (ms): 3.1 3.2 0.2 0.0 1.0 1.3 17.2 0.2 0.0 0.9 1.1 3.1 1.7 0.0 1.4 1.1 8.7 0.6 1.0 Global cache hit ratio: Ratio of current block defers: % of messages sent for bu ffer gets: % of remote buffer gets: Ratio of I/O for coherence: Ratio of local v s remote work: Ratio of fusion vs physical writes: In the above report, the statistics in relation to the other instances in the cl uster should be examined. The possible causes of any statistics that are not in line with the other cluster instances should be investigated By making the netwo rk changes . and index changes stated before, the workload was increased by a fa ctor of greater than three, and the response time was still less than in the ori ginal STATSPACK. The following is the same section from the STATSPACK report net work changes. Almost all statistics show an increase: Cluster Statistics for DB: MIKE Instance: mike2 Snaps: 105 -106 taken after the Global Cache Service - Workload Characteristics ---------------------------------------------Ave global cache get time (ms): 8.2
Ave global cache convert time (ms): Ave Ave Ave Ave Ave Ave Ave Ave Ave Ave buil d time for CR block (ms): flush time for CR block (ms): send time for CR block ( ms): time to process CR block request (ms): receive time for CR block (ms): pin time for current block (ms): flush time for current block (ms): send time for cu rrent block (ms): time to process current block request (ms): receive time for c urrent block (ms): 16.5 1.5 6.0 0.9 8.5 18.3 13.7 3.9 0.8 18.4 17.4 2.5 0.2 2.2 1.6 2.8 0.5 0.0 Global cache hit ratio: Ratio of current block defers: % of messages sent for bu ffer gets: % of remote buffer gets: Ratio of I/O for coherence: Ratio of local v s remote work: Ratio of fusion vs physical writes: The next report shows the global enqueue service statistics. The global enqueue services (GES) con trol the inter -instance locks in Oracle Database 10g RAC. These times should all be in the less than 15 millisecond range and the ratio sh ould be near one. If they are not, it shows possible network or memory problems. Application locking issues can also ause this and the enqueue report, which is c shown later in the STATSPACK listing, should be consulted to further diagnose the possible problems: Global Enqueue Service Statistics -------------------------------Ave global lock get time (ms): Ave global lock convert time (ms): Ratio of global lock gets vs global lock releases: 0.9 1.3 1.1 The next STATSPACK report deals with GCS and GES messaging. Queue times greater than 20-30 ms should be consi dered excessive and should be watched: GCS and GES Messaging statistics -------------------------------Ave message sent queue time (ms): 1.8 Ave message sent queue time on ksxp (ms): 2.6 Ave message received qu eue time (ms): 1.2 Ave GCS message process time (ms): 1.2 Ave GES message proces s time (ms): 0.2 % of direct sent messages: 58.4 % of indirect sent messages: 4. 9 % of flow controlled messages: 36.7 -----------------------------------------------------------The next section of the STATSPACK report also deals with the global enqueue serv ices statistic Blocked converts are one thing to look for. There are several in s. this report example. Blocked converts means the instance requested a block fr om another instance and for one reason or another was unable to obtain the conversion of the block. This can indicate that users are going after the same r ecords, and it may be desirable to prevent that from occurring.
freelists, ini_trans, and limiting rows per block to avoid conversions being blo cked. If there are excessive message processing times, thought should be given t o tuning the network to increase bandwidth, or perhaps upgradi ng the NIC s to f aster, high bandwidth versions. GES Statistics for DB: MIKE Instance: mike2 Snaps: 25 -26 Statistic Total per Se cond per Trans --------------------------------- ---------------- ------------ ----------dynamically allocated gcs resourc 0 0.0 0.0 dynamically allocated gcs shadows 0 0.0 0.0 flow control messages received 0 0.0 0.0 flow control messages sent 0 0.0 0.0 gcs ast xid 0 0.0 0.0 gcs blocked converts 904 2.1 150.7 gcs blo cked cr converts 1,284 2.9 214.0 gcs compatible basts 0 0.0 0.0 gcs compatible c r basts (global) 0 0.0 0.0 gcs compatible cr basts (local) 75 0.2 12.5 gcs cr ba sts to PIs 0 0.0 0.0 gcs cr serve without current lock 0 0.0 0.0 gcs error msgs 0 0.0 0.0 gcs flush pi msgs 21 0.0 3.5 gcs forward cr to pinged instance 0 0.0 0 .0 gcs immediate (compatible) conver 4 0.0 0.7 gcs immediate (null) converts 79 0.2 13.2 gcs immediate cr (compatible) con 4 0.0 0.7 gcs immediate cr (null) con verts 3 0.0 0.5 gcs msgs process time(ms) 3,193 7.3 532.2 gcs msgs received 2,58 6 5.9 431.0 gcs out-of-order msgs 0 0.0 0.0 gcs pings refused 0 0.0 0.0 gcs queu ed converts 0 0.0 0.0 gcs recovery claim msgs 0 0.0 0.0 gcs refuse xid 0 0.0 0.0 gcs retry convert request 0 0.0 0.0 gcs side channel msgs actual 65 0.1 10.8 gc s side channel msgs logical 1,383 3.2 230.5 gcs write notification msgs 0 0.0 0. 0 gcs write request msgs 0 0.0 0.0 gcs writes refused 0 0.0 0.0 ges msgs process time(ms) 136 0.3 22.7 ges msgs received 578 1.3 96.3 implicit batch messages re ceived 90 0.2 15.0 implicit batch messages sent 12 0.0 2.0 lmd msg send time(ms) 55 0.1 9.2 lms(s) msg send time(ms) 8 0.0 1.3 messages flow controlled 806 1.8 134.3 messages received actual 1,980 4.5 330.0 messages received logical 3,164 7 .3 527.3 messages sent directly 1,281 2.9 213.5 messages sent indirectly 108 0.2 18.0 msgs cau sing lmd to send msgs 231 0.5 38.5 msgs causing lms(s) to send ms gs 97 0.2 16.2 msgs received queue time (ms) 3,842 8.8 640.3 msgs received queue d 3,164 7.3 527.3 msgs sent queue time (ms) 202 0.5 33.7 msgs sent queue time on ksxp (ms) 4,337 9.9 722.8 msgs sent queued 111 0.3 18.5 msgs sent queued on ksx p 1,658 3.8 276.3 process batch messages received 269 0.6 44.8 process batch mes sages sent 191 0.4 31.8 -----------------------------------------------------------The next section of the report deals with waits. Waits are a key tuning indicato r. global cache cr request, which was caused by t he network not being tuned proper ly, as already mentioned. The second highest wait is the global cache null to x, which, if severe, indicates problems with the I/O subsystem. In this case, the total time waited was one second or less, hardly a concern when compared with th e 154 second wait on global cache cr request. The waits with the highest total t ime should be tuned first. Waits such as SQL*Net waits and any The predominant wait is for
smon, pmon, or wakeup timers can be safely ignored in most situations. Wait Events for DB: MIKE Instance: mike2 Snaps: 25 -26 -> s - second -> cs - cen tisecond 100th of a second -> ms - millisecond 1000th of a second -> us - micros econd - 1000000th of a second -> ordered by wait time desc, waits desc (idle eve nts last) Avg Total Wait wait Waits Event Waits Timeouts Time (s) (ms) /txn --------------------------- ------------ ---------- ---------- ------ -------global cache cr request 820 113 154 188 136.7 global cache null to x 478 1 1 2 79.7 co ntrol file sequential read 600 0 1 2 100.0 control file parallel write 141 0 1 4 23.5 enqueue 29 0 1 18 4.8 library cache lock 215 0 0 2 35.8 db file sequential read 28 0 0 7 4.7 LGWR wait for redo copy 31 16 0 4 5.2 ksxr poll remote instan ces 697 465 0 0 116.2 global cache open x 48 0 0 2 8.0 CGS wait for IPC msg 899 899 0 0 149.8 log file parallel write 698 697 0 0 116.3 latch free 24 0 0 2 4.0 global cache s to x 41 0 0 1 6.8 log file sync 3 0 0 13 0.5 DFS lock handle 30 0 0 1 5.0 global cache open s 16 0 0 1 2.7 global cache null to s 9 0 0 1 1.5 lib rary cache pin 133 0 0 0 22.2 KJC: Wait for msg sends to c 17 0 0 0 2.8 db file parallel write 19 0 0 0 3.2 cr request retry 27 27 0 0 4.5 gcs remote message 10 ,765 9,354 840 78 1,794.2 virtual circuit status 15 15 440 29307 2.5 ges remote message 9,262 8,501 421 45 1,543.7 wakeup time manager 14 14 408 29135 2.3 SQL*N et message from client 4,040 0 216 53 673.3 SQL*Net message to client 4,040 0 0 0 673.3 ------------------------------------------------------------The next report deals with enqueues. non level enqueue These are the normal system enqueues. The are high the -RAC related ones have been removed from this listing. Enqueues locks used to pr otect memory areas. When the report is reviewed, s with the highest totals should be of the most interest. In the following report, all of the enqueues of concern are again dealing with message times and cache block transfers. These types of enqueues again point to network tuning. Enqueue activity for DB: MIKE Instance: mike2 Snaps: 25 -26 -> Enqueue stats gat hered prior to 10g should not be compared with 10g data -> ordered by Wait Time desc, Waits desc Instance Activity Stats for DB: MIKE Instance: mike2 Snaps: 25 -26 Statistic Total per Second per Trans --------------------------------- ----------------- -------------- -----------gcs messages sent 1,570 3.6 261.7 ges messag es sent 805 1.9 134.2 global cache blocks lost 88 0.2 14.7 global cache convert time 171 0.4 28.5 global cache converts 528 1.2 88.0 global cache cr block build time 28 0.1 4.7 global cache cr block flush time 2 0.0 0.3 global cache cr bloc k receive tim 1,158 2.7 193.0 global cache cr block send time 117 0.3 19.5 globa l cache cr blocks received 674 1.6 112.3 global cache cr blocks served 1,147 2.6 191.2 global cache current block pin ti 12 0.0 2.0 global cache current block r eceiv 170 0.4 28.3 global cache current block send t 57 0.1 9.5 global cache cur rent blocks recei 541 1.2 90.2 global cache current blocks serve 653 1.5 108.8 g lobal cache defers 0 0.0 0.0 global cache get time 57 0.1 9.5 global cache gets 183 0.4 30.5 global cache skip prepare failure 37 0.1 6.2 global lock async conv erts 0 0.0 0.0 global lock async gets 197 0.5 32.8 global lock convert time 4 0. 0 0.7
global global global global lock lock lock lock get time releases sync converts sync gets 290 3,064 30 3,120 0.7 7.0 0.1 7.2 48.3 510.7 5.0 520.0 The next three reports deal with latches. Latches are low level lock structures that protect memory areas. There is usually no concern with latches unless high sleeps or misses are observed. Generally, adjusting internal parameters or tun n g SQL i _kgl* parameters to tune the library related or shared pool locks. Sometimes, in creasing the shared pool size can also help to relieve latch issues. For some la tches, the number of latches avai lable are derived from the size of the shared pool and the settings for related initialization parameters. code tunes latches. An example would be adjusting the Latch Activity for DB: MIKE Instance: mike2 Snaps: 25 -26 ->"Get Requests", "Pct Get Miss" and "Avg Slps/Miss" are statistics for willing-to-wait latch get requ ests ->"NoWait Requests", "Pct NoWait Miss" are for no -wait latch get requests ->"Pct Misses" for both should be very close to 0.0 Pct Avg Wait Pct Get Get Slp s Time NoWait NoWait Latch Requests Miss /Miss (s) Requests Miss ----------------------- -------------- ------ ------ ------ ------------ -----KCL bast context freelis 1,150 0.0 0 0 KCL freelist parent latc 713 0.0 0 0 KCL gc element paren t la 8,399 0.0 0 166 0.0 KCL name table parent la 1,780 0.0 0 144 0.0 KJC messag e pool free li 655 0.0 0 194 0.0 KJCT flow control latch 2,238 0.0 1.0 0 91 0.0 gcs opaque info freelist 1,922 0.0 0 0 gcs resource freelist 77 0.0 0 0 gcs reso urce hash 5,719 0.0 1.0 0 0 gcs shadows freelist 681 0.0 0 0 ges caches resource list 3,518 0.0 0 2,130 0.0 ges deadlock list 411 0.0 0 0 ges domain table 4,206 0.0 0 0 ges enqueue table freeli 6,271 0.0 0 0 ges group parent 4,238 0.0 0 0 g es group table 6,409 0.0 0 0 ges process hash list 207 0.0 0 0 ges process paren t latch 18,354 0.0 1.0 0 0 ges process table freeli 4 0.0 0 0 ges resource hash list 10,294 0.1 1.0 0 1,248 0.0 ges resource table freel 5,703 0.0 0 0 ges timeo ut list 72 0.0 0 0 Latch Sleep breakdown for DB: MIKE -> ordered by misses desc Instance: mike2 Snaps: 25 -26 Get Spin & Latch Name Requests Misses Sleeps Sleeps 1 ->4 ------------------------- -------------- ----------- ----------- -----------ges resource hash list 10 ,294 13 13 0/13/0/0/0 redo allocation 41,394 4 4 0/4/0/0/0 KJCT flow control lat ch 2,238 1 1 0/1/0/0/0 cache buffers lru chain 3,055 1 1 0/1/0/0/0 gcs resource hash 5,719 1 1 0/1/0/0/0 ges process parent latch 18,354 1 1 0/1/0/0/0 -----------------------------------------------------------Latch Miss Sources for DB: MI KE Instance: mike2 Snaps: 25 -26 -> only latches with sleeps are shown -> ordere d by name, sleeps desc NoWait Waiter Latch Name Where Misses Sleeps Sleeps ----------------------- -------------------------- ------- ---------- -------KJCT fl ow control latch kjcts_sedeqv: dequeue a ve 0 1 0 cache buffers lru chain kcbzgb : wait 0 1 1 gcs resource hash kjbmpconvert 0 1 1 ges resource hash list kjrmas1 : lookup master nod 0 13 0 redo allocation kcrfwr 0 4 2 -----------------------------------------------------------The next report deals with the data dictionary cache area of the shared pool. If there are no conflicts or releases, the dc caches are sized correctly. The dictionary cache area is sized by properly sizin g the shared pool.
Dictionary Cache Stats for DB: MIKE Instance: mike2 GES GES Snaps: 25 -26 GES
Cache Requests Conflicts Releases ------------------------- ------------ ----------- -----------dc_global_oids 0 0 0 dc_object_ids 22 0 0 dc_objects 97 0 0 dc_ profiles 0 0 0 dc_rollback_segments 0 0 0 dc_segments 3 0 0 dc_tablespace_quotas 2 0 0 dc_tablespaces 0 0 0 dc_user_grants 0 0 0 dc_usernames 0 0 0 dc_users 0 0 0 ------------------------------------------------------------The nex t report gives a breakdown of enqueues by type of enqueue. This enqueue type report will help isolate the cause of enqueue problems. The types of enqueue are shown in Ta ble 14.2 Type BL CF CI CU DF DL DM DR DX FS IN IR IS IV JQ KK L[A-P] MR N[A-Z] PF PI PR P S Q[A-Z] RT SC SM SQ SR SS ST SV Description Buffer Cache Management Controlfile Transaction Cross-instance Call Invocation Bind Enqueue Datafile Direct Loader Index Creation Database Mount Dis tributed Recovery Distributed TX File Set Instance Number Instance Recovery Inst ance State Library Cache Invalidation Job Queue Redo Log "Kick" Library Cache Lo ck Media Recovery Library Cache Pin Password File Parallel Slaves Process Startu p Parallel Slave Synchronization Row Cache Redo Thread System Commit Number SMON Sequence Number Enqueue Synchronized Replication Sort Segment Space Management Transaction Sequence Number Value
Type TA TM TS TT TX UL UN US WL XA XI Table 14.2: Description Transaction Recovery DML Enqueue Temporary Segment (also TableSpace) Temporary Table Transaction User-defined Locks User Name Undo Segment: Serializ ation Being Written Redo Log Instance Attribute Lock Instance Registration Lock List of Enqueues In the report, the majority enqueue is the TM or DML relate d enqueue. However, its average wait time is only 2.43 milliseconds. The transaction recovery (TA) e nqueue has a whopping 497 millisecond wait time. This high value was driven by s everal rollbacks in the test procedure. Once the enqueue causing the problem s a re determined, standard tuning techniques can be used to resolve them. Enqueue activity for DB: MIKE Instance: mike2 Snaps: 25 -26 -> Enqueue stats gat hered prior to 10g should not be compared with 10g data -> ordered by Wait Time desc, Waits desc Avg Wt Wait Eq Requests Succ Gets Failed Gets Waits Time (ms) T ime (s) -- ------------ ------------ ----------- ----------- ------------- ----------TA 1 1 0 1 497.00 0 TM 1,582 1,582 0 14 2.43 0 HW 13 13 0 5 2.60 0 FB 4 4 0 4 3.00 0 TT 3 3 0 3 2.33 0 -----------------------------------------------------------The final report section to be reviewed is the library cache re GES. GES invalid requests and GES invalidations could indicate insufficient port dealing with the sizing of the shared pool, resulting in GES contention. Library Cache Activity for DB: MIKE ->"Pct Misses" should be very low Instance: mike2 Snaps: 25 -26 GES Lock GES Pin GES Pin GES Inval GES InvaliNamespace Requests Requests Release s Requests dations --------------- ------------ ------------ ------------ ---------- ----------BODY 1 0 0 0 0 CLUSTER 4 0 0 0 0 INDEX 84 0 0 0 0 SQL AREA 0 0 0 0 0 TABLE/PROCEDURE 617 192 0 77 0 TRIGGER 0 0 0 0 0 -----------------------------------------------------------This section on the STATSPACK reports has only covered the reports that dealt wi th the RAC environment. This does not mean that the rest of the report can be ig nored. The sections on SQL use are critical in light of the need to find the tha t may actually be causing the problems. The other waits, latches, enqueues, and code
global cache gets + global cache converts + global cache cr blocks received + gl obal cache current blocks received) by the number of logical reads ( consistent gets + db block gets) for a given stati stics collection interval. A global cach e service request is made in Oracle when a user attempts to access a buffer cach e to read or modify a data block and the block is not in the local cache. A remo te cache read, disk read or change access privileges is the inevitable result. T hese are logical read related. Logical reads form a superset of the global cache service operations. The calculation for global cache hit ratio since instance s tartup is: SELECT a.inst_id "Instance", (A.VALUE+B.VALUE+C.VALUE+D.VALUE)/(E.VALUE+F.VALUE) "GLOBAL CACHE HIT RATIO" FROM GV$SYSSTAT A, GV$SYSSTAT B, GV$SYSSTAT C, GV$SYSS TAT D, GV$SYSSTAT E, GV$SYSSTAT F WHERE A.NAME= gc gets AND B.NAME= gc converts AND C.NAME= gc cr blocks received AND D.NAME= gc current blocks received AND E.NAME= consistent gets AND F.NAME= db block gets SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.oracle-script.com/) Instance GLOBAL CACHE HIT RATIO ---------- ---------------------1 .02403656 2 .0 14798887 The instance with the best access to the drives, or the faster I/O path, will li kely cache hit ratio. This is due to the way Oracle s RAC caching algorithm works as it may decide that the cost of doing a local read is higher than reading into th e other cache and siphoning it across the cluster interconnect. In formula form: have the best (gc gets + gc converts + gc cr blocks received + gc current blocks received) / (consistent gets + db block gets) Blocks frequently requested by local and remote users will be very hot. If a blo ck is hot, its transfer their is delayed The for a few ratio milliseconds provid es a to allow the local us to ers of how complete work. following rough estimate prevalent this is: SELECT
A.INST_ID "Instance", A.VALUE/B.VALUE "BLOCK TRANSFER RATIO" FROM GV$SYSSTAT A, GV$SYSSTAT B WHERE A.NAME= gc defers AND B.NAME= gc current blocks served SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.oracle-script.com/) Instance BLOCK TR ANSFER RATIO ---------- -------------------1 .052600105 2 .078004479 If the above SELECT generates a ratio of more than 0.3, a fairly hot data set is indicated. If this is the case, blocks involved in busy waits should be analyze d. The following columns should be queried to find the blocks involved in busy w aits: n n n n name kind forced_reads forced_writes For example: col instance format 99999999 col name format a20 col kind format a10 set lines 8 0 pages 55 Select INST_ID "Instance", NAME, KIND, sum(FORCED_READS) "Forced Read s", sum(FORCED_WRITES) "Forced Writes" FROM GV$CACHE_TRANSFER SEE GRID CODE DEPO T FOR DOWNLOAD (http://www.rampantbooks.com/book_2004_1_10g_grid.htm) Instance N AME KIND Forced Reads Forced Writes --------- -------------------- ---------- ----------- ------------1 MOD_TEST_IND INDEX 308 0 1 TEST2 TABLE 64 0 1 AQ$_QUEUE _TABLES TABLE 5 0 2 TEST2 TABLE 473 0 2 MOD_TEST_IND INDEX 221 0 2 AQ$_QUEUE_TAB LES TABLE 2 0 gv$cache_transfer view. Alternatively, the mns in gv$cr_block_server can be examined. cache busy, buffer busy global cache, and the gv$sysstat view should be examined. These values come from the SELECT INST_ID, URRENT_REQUESTS) "Current Requests" FROM cr_requests and current_requests colu Also, the values shown for the global buffer busy global cr statistics from sum(CR_REQUESTS) "CR Requests", sum(C
GV$CR_BLOCK_SERVER GROUP BY INST_ID; INST_ID CR Requests Current Requests --------- ----------- ---------------1 28940 2244 2 31699 837 SELECT inst_id "Instanc e", event "Wait Event", total_waits, time_waited FROM GV$SYSTEM_EVENT WHERE even t in ( global cache busy , buffer busy global cache , buffer busy global CR ) SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.oracle-script.com/) Instance -------1 1 2 Wait Event TOTAL_WAITS TIME_WAITED ------------------------ ---------- ----------buffer busy global CR 1 0 global cache busy 1073 7171 global cache busy 973 7524 If a problem is discovered, the object causing the problem should be identified along with the instance that is accessing the object, and how the object is bein g accessed. If necessary, the contention an be alleviated by: c n n Reducing hot spots by spreading the accesses to index blocks or data blocks. Using Oracle hash or range partitions wherever applicable, just as it would be d one in single instance Oracle databases. n Reducing concurrency on the object by implementing load balancing or resource management. For example, decrease the rate of modifications to that obj ect by using fewer database processes. In RAC, as in a single instance Oracle da tabase, blocks are only written to disk for aging, cache replace ment, or checkp oints. When a data block is replaced from the cache due to aging or when a checkpoint occurs and the block was previously chan ged in another instance but not written to disk, Oracle sends a message to notif y the other instance that Oracle wi block to disk. ll perform a fusion write to move the data These fusion writes a re monitored with the following
ratio. It reveals the proportion of writes that Oracle manages. SELECT a.inst_id "Instance", A.VALUE/B.VALUE "Cache Fusion Writes Ratio" FROM GV $SYSSTAT A, GV$SYSSTAT B WHERE
a.name= DBWR fusion writes AND b.name= physical writes AND b.inst_id=a.inst_id ORDER BY A.INST_ID; Instance Cache Fusion Writes Ratio --------- -----------------------1 .216290958 2 .131862042 The larger this ratio is, the higher the number of written blocks that have been copied with their previous changes between the RAC instances. A large ratio is the result of: n n n Insufficiently sized caches. Insufficie nt checkpoints. Large numbers of buffers written due to cache replacement or checkpointing. For example, 0.21 means that 21% of the buffers written to disk were globally di rty. A fusion write does not involve an additional write to disk. A fusion write do es require messaging to arrange the transfer with the other instances. This indicates that fusion writes are in fact a subset of all the instances physical w rites. Use of the v$ cache_transfer Views The v$cache_transfer and v$file_cache_transfer views are used to examine RAC statistics. The types of blocks that use the cluster interconnects in a RAC envi ronment are monitored with the v$ cache transfer series of views: n v$cache_transfer: This view shows the types and classes of blocks that Oracle tr ansfers over the cluster interconnect on a per -object basis. The forced_reads a nd forced_writes columns can be used to determine the types of objects the RAC i nstances are sharing. Valu es in the forced_writes column show how often a certa in block type is transferred out of a local buffer cache due to the current vers ion being requested by another instance. The following columns are the same for all of the views. v$class_cache_transfer: -- This view can be used to identify t he class of blocks that experience cache transfers. v$class_cache_transfer has a class column showing the class of a block; therefore, this view can be used to assess block transfers per c lass of object. v$file_cache_transfer: This view ca n be used to monitor the blocks transferred per file. The file_number column ide ntifies the datafile that contained the blocks transferred. v$temp_cache_transfe r: -- This view can be used to monitor the transfer of temporary tablespace bloc ks. The view contains a file_number column that is n n n
v$temp_cache_transfer view. The contents of the v$cache_transfer view are shown below. Description of the V$CACHE_TRANSFER view Name ----------------------FILE# BLOCK# CLASS# STATUS XNC FORCED_READS FORCED_WRITES NAME PARTITION_NAME KIND OWNER# GC _ELEMENT_ADDR GC_ELEMENT_NAME Type -------------NUMBER NUMBER NUMBER VARCHAR2(5) NUMBER NUMBER NUMBER VARCHAR2(30) VARCHAR2(30) VARCHAR2(15) NUMBER RAW(4) NUMBE R v$cache_transfer view shows the types and classes of blocks that Oracle transfer s over the cluster interconnect on a per -object basis. The forced_reads and for ced_writes columns are used to determine which types of objects the RAC instance s are sharing. The The v$file_cache_transfer view is used to identify files that have experienced c ache transfers. For example, while v$cache_transfer has a name column showing th e name of an object, the v$file_cache_transfer view has the file_number co lumn to show the file numbers of the datafiles that are the source of blocks transfer red; therefore, this view can be used to assess block transfers per file. The v$ file_cache_transfer view contents are shown below: Description of the V$FILE_CACHE_TRANSFER View Name Type ------------------------------- ---------FILE_NUMBER NUMBER X_2_NULL NUMBER X_2_NULL_FORCED_WRITE NUMBE R X_2_NULL_FORCED_STALE NUMBER X_2_S NUMBER X_2_S_FORCED_WRITE NUMBER S_2_NULL N UMBER S_2_NULL_FORCED_STALE NUMBER RBR NUMBER RBR_FORCED_WRITE NUMBER RBR_FORCED _STALE NUMBER NULL_2_X NUMBER S_2_X NUMBER NULL_2_S NUMBER CR_TRANSFERS NUMBER C UR_TRANSFERS NUMBER
v$cache_transfer and v$file_cache_transfer views may still show the number of bl ock mode conversions per block class or object. Values in the forced_writes colu mn, however, will be zero. Monitoring the GES Processes The monitoring of the gl obal enqueue services (GES) process is performed using the gv$enqueue_stat view. The contents of the gv$enqueue_stat view are s hown below: Description of the view GV$ENQUEUE_STAT Name -------------------------------INST _ID EQ_TYPE TOTAL_REQ# TOTAL_WAIT# SUCC_REQ# FAILED_REQ# CUM_WAIT_TIME Type --------NUMBER VARCHAR2(2) NUMBER NUMBER NUMBER NUMBER NUMBER An example SELECT to retrieve all of the enqueues with total_wait number greater than zero would be: select * from gv$enqueue_stat where total_wait#>0 order by inst_id, cum_wait_tim e desc; SEE CODE DEPOT FOR COMPLETE SCRIPT (http://www.oracle-script.com/) INST_ID ---------1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 EQ TOTAL_ REQ# TOTAL_WAIT# SUCC_REQ# FAILED_REQ# CUM_WAIT_TIME -- ---------- ----------- --------- ----------- ------------TX 31928 26 31928 0 293303 PS 995 571 994 1 55 658 TA 1067 874 1067 0 10466 TD 974 974 974 0 2980 DR 176 176 176 0 406 US 190 1 89 190 0 404 PI 47 27 47 0 104 CF 499314 23 499314 0 47 TM 41928 8 41928 0 35 MR 93 13 93 0 21 HW 637 6 637 0 8 XR 4 2 4 0 5 DM 4 2 4 0 4 SR 1 1 1 0 3 SW 3 1 3 0 3 TS 2 1 2 0 3 TA 1064 1015 1064 0 437648 PS 2208 325 1597 611 104273 TX 44084 3 18 440843 0 62787 US 197 41 197 0 8551 IR 193 29 193 0 4593 TT 4393 131 4393 0 3363 CF 507497 540 507497 0 1726 TM 1104694 101 1104694 0 766 DM 5 2 5 0 483 HW 444 41 444 0 108 PI 90 18 90 0 81
2 2 2 2 2 2 2 DL DR RT FB IA PG TS 32 1 76 4 10 1 1 4 18 23 4 2 1 1 1 32 176 3 10 1 1 4 0 0 1 0 0 0 0 55 52 12 6 4 3 3 According to Oracle, the enqueues of interest, as shown in the the gv$enqueue_stat view in the RAC environment are: n SQ Enqueue : This indicates that there is contention for sequences. In almost all cases, executing an ALTER SEQUENCE command can increase the cache size of se quences used by the applicatio environment, DBAs should use the NOORDER keyword to eq_type column of n. When creating sequences for a RAC avoid an additional cause of SQ enqueue contention that is forced ordering of queued sequ ence values. n : This is usually an application related issue pertaini ng to row locking. Real Application Clusters processing can magnify the effect of TX enqueue waits. Perf ormance bottlenecks can also appear on leaf blocks of right growing indexes as T X enqueue waits while the index block splits are occuring. TX enqueue p erforman ce issues can be resolved by setting the value of the initrans parameter for a T ABLE or INDEX to be equal to the number of CPUs per node multiplied by the numbe r of nodes in the cluster multiplied by 0.75. Another technique is to d etermine the number of simultaneous accesses for DML for the objects experiencing TX enq ueues, and setting initrans to that value. Oracle Corporation recommends avoidin g setting this parameter greater than 100. Another parameter that can reduce TX enqueues is maxtrans. maxtrans determines the maximum number of transactions tha t can access a block. maxtrans will default to 255 and it is a good practice to reset this to less than 100. TX Enqueue PS (Parallel Slave Synchronization) and TA (Transaction Re covery) enqueues also seem to have some importance in the environment. Therefore, start with a wide s weep and then focus on the waits that are causing performance issues in the envi ronment. There are other interesting views that provide information on tun in a RAC environment should be aware of, for example: n ing that the DBA gv$segment_statistics: Provides statistics such as buffer busy waits on a per se gment basis. This allows tracking of exactly whi ch segments, indexes or tables, are causing the buffer busy waits or other statistics to increment. To select a gainst the gv$segment_statistics view, the user will want to SELECT for a specif ic statistic name where the value is greater than a predetermined limit. The con tents of gv$segment_statistics are shown below: Description of gv$segment_statistics Name Null? Type ---------------------------
INST_ID OWNER OBJECT_NAME SUBOBJECT_NAME TABLESPACE_NAME TS# OBJ# DATAOBJ# OBJEC T_TYPE STATISTIC_NAME STATISTIC# VALUE NUMBER VARCHAR2(30) VARCHAR2(30) VARCHAR2(30) VARCHAR2(30) NUMBER NUMBER NUMBER VARCHAR2(18) VARCHAR2(64) NUMBER NUMBER Another useful view shows which file IDs which have been remastered, which happens when they are transferred from one instance to another, this view is cal led gv$gcspfmaster_info and its contents are shown below: Description of gv$gcspfmaster_info Name Null? ---------------------------------------- -------INST_ID FILE_ID CURRENT_MASTER PREVIOUS_MASTER REMASTER_CNT Type ------NUMBER NUMBER NUMBER NUMBER NUMBER file_id column corresponds to the data file ID. The current and previous masters refer to the instances that are either the curren t or previous master of the s pecified file. A view related to the gv$gcspfmaster_info view is the gv$gchvmast er_info view which shows the same information but for the PCM hash value IDs for specific r esources that have been remastered. This views contents are shown be low: The Description of gv$gcshvmaster_info Name Null? ---------------------------------------- -------INST_ID HV_ID CURRENT_MASTER PREVIOUS_MASTER REMASTER_CNT Typ e ------NUMBER NUMBER NUMBER NUMBER NUMBER To select against these views, it may be desirable to restrict on the remaster_c nt value being greater than a predetermined limit. The gv$sqlarea view has also bee n enhanced in Oracle Database 10g RAC. The column cluster_wait_time in gv$sqlare a represents the wait time incurred by individual SQL statements for global cach e events and will identify the SQL which may need to be tuned based on its contr ibution to RAC contention. Monitoring and Tuning using OEM The first part of thi s chapter covered the manual aspects of RAC tuning, showing techniques and provi ding scripts to obtain the needed information from the various underlying Oracle views. This section will cover the tools provided by
n n n OEM installed. OEM repository. Oracle Intelligent Agents running. Normally, if the standard installation is used, OEM will be installed as a matt course. If possible, the OEM and its repository should be set up on a separate s ystem, such as a small Windows - based server. At the least, a small repository er of database should be created on one of the RAC servers. This can be used for the R MAN reposito ry as well. The OEM client software can be installed on any desktop as long as that desktop has SQL*Net connectivity to the OEM repository. The most difficult part of the setup for OEM is probably getting the intelligent agents to work properly with the RAC environment. Configuring the Oracle Intelligent Agent with RAC In order to be sure that the agent is installed properly in Oracle Database 10g, the following Metalink documents should be reviewed: n n n n EM 10g Database Co ntrol Release Notes 10.1.0.2.0 Note: 266770.1 EM 10g GRID Control Release Notes 10.1.0.2.0 Note: How to Log and Trace the EM 1 0g Management Agents Note: 266769.1 229624.1 How To Install The Downloadable Central Management Agent in EM 10g Grid Control Note: 235287.1 The nodes and databases mus OEM webpage before they can be monitored. t be properly discovered and viewable from the Using EM in Oracle Database 10g In 10g, Oracle is moving to the web based interface for EM and will soon depreca te the Java based version. Monitoring and management fu be migrated into the HTM L Web based version of EM as soon as possible. The following sections will provi de a quick look at the screens in the new HTML version that the DBA should be us ing for RAC monitoring and management. nctions should The Cluster Perfor mance Page In the HTML based web version of the EM, the Cluster Performance Page is used to display the usage statistics for all RAC hosts or for individual RAC hosts. Thi s
Example Cluster Performance Page The Cluster Database Performance Page The Oracl e Database 10g EM Cluster Database Performance Page displays statistics via char ts that show run queue length, paging rate, serv ice time, and the database thro ughput for each RAC host or RAC instance. The page is also used to access the de tailed information for the Wait Class Page for Service Time and the Top Sessions Page for Database Throughput. Figure 14. 2 shows an example of this page.