Prerequisites for awrreport. Statistics level should be set to typical. Timed statistics should also be enabled. SQL> show parameter statistics_levelNAME TYPE VALUE ------------------------------------ -------- ------------------------------ statistics_level string TYPICAL
SQL> select STATISTICS_NAME,ACTIVATION_LEVEL, SYSTEM_STATUS from v$statistics_level;STATISTICS_NAME ACTIVAT SYSTEM_S ---------------------------------------------------------------- ------- -------- Buffer Cache Advice TYPICAL ENABLED MTTR Advice TYPICAL ENABLED Timed Statistics TYPICAL ENABLED Timed OS Statistics ALL DISABLED Segment Level Statistics TYPICAL ENABLED PGA Advice TYPICAL ENABLED Plan Execution Statistics ALL DISABLED Shared Pool Advice TYPICAL ENABLED Modification Monitoring TYPICAL ENABLED Longops Statistics TYPICAL ENABLED Bind Data Capture TYPICAL ENABLED Ultrafast Latch Statistics TYPICAL ENABLED Threshold-based Alerts TYPICAL ENABLED Global Cache Statistics TYPICAL ENABLED Active Session History TYPICAL ENABLED Undo Advisor, Alerts and Fast Ramp up TYPICAL ENABLED Streams Pool Advice TYPICAL ENABLED Time Model Events TYPICAL ENABLED Plan Execution Sampling TYPICAL ENABLED Automated Maintenance Tasks TYPICAL ENABLED SQL Monitoring TYPICAL ENABLED Adaptive Thresholds Enabled TYPICAL ENABLED V$IOSTAT_* statistics TYPICAL ENABLED
The Header The header basically gives you a Bird’s-eye view information on the database.WORKLOAD REPOSITORY report for DB Name DB Id Instance Inst Num Release RAC Host ------------ ----------- ------------ -------- ----------- --- ------------ DDOG 1373267642 DDOG2 2 10.2.0.2.0 YES lefty Snap Id Snap Time Sessions Curs/Sess --------- ------------------- -------- --------- Begin Snap: 2009 09-Jan-07 09:00:12 40 3.7 End Snap: 2010 09-Jan-07 10:00:19 39 2.8 Elapsed: 60.12 (mins) DB Time: 0.09 (mins) Cache Sizes ~~~~~~~~~~~ Begin End ---------- ---------- Buffer Cache: 528M 528M Std Block Size: 8K Shared Pool Size: 480M 480M Log Buffer: 15,152K
Elasped Time: It represents the snapshot window or the time between the two snapshots. DB TIME: Represents the activity on the database. If DB TIME is Greater than Elapsed Time then it means that database has high workload. Load Profile Meaning of various terms. •Redo size – Indicates the amount of DML activity happening in the database. •Logical and physical reads – Represents number of IO's (Physical and logical) that the database is performing. •User calls - Indicates how many user calls have occurred during the snapshot period. This value can give you some indication if usage has increased. •Parses and hard parses - Provides an indication of the efficiency of SQL re-usage. •Sorts - Number of sorts occurring in the database. •Logons – Number of logins which occurred in the database. •Executes – Number of SQL statements which were executed. •Transactions - Indicates how many transactions occurred during the snapshot period.Load Profile ~~~~~~~~~~~~ Per Second Per Transaction --------------- --------------- Redo size: 209.78 1,014.29 Logical reads: 20.14 97.38 Block changes: 0.50 2.42 Physical reads: 0.00 0.01 Physical writes: 0.10 0.49 User calls: 1.05 5.08 Parses: 0.39 1.86 Hard parses: 0.00 0.01 Sorts: 0.52 2.53 Logons: 0.02 0.12 Executes: 0.95 4.61 Transactions: 0.21 % Blocks changed per Read: 2.49 Recursive Call %: 84.53 Rollback per transaction %: 97.45 Rows per Sort: 41.43
The percentage of soft parses should always be higher than hard parses. Possible reasons for excessive hard parses may be a small shared pool Or may be that bind variables are not being used. Instance EfficiencyInstance Efficiency Percentages (Target 100%) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Buffer Nowait %: 99.98 Redo NoWait %: 100.00 Buffer Hit %: 100.00 In-memory Sort %: 100.00 Library Hit %: 99.64 Soft Parse %: 99.64 Execute to Parse %: 59.60 Latch Hit %: 100.00 Parse CPU to Parse Elapsd %: 130.00 % Non-Parse CPU: 96.30 Shared Pool Statistics Begin End ------ ------ Memory Usage %: 87.98 87.97 % SQL with executions>1: 97.83 97.86 % Memory for SQL w/exec>1: 85.54 85.52
The following ratios should be above 90% in a database. Buffer Nowait Buffer Hit Library Hit Redo NoWait In-memory Sort Soft Parse Latch Hit Non-Parse CPU The execute to parse ratio should be very high in a ideal database. The execute to parse ratio is basically a measure between the number Of times a sql is executed versus the number of times it is parsed. The ratio will move higher as the number of executes go up, while The number of parses either go down or remain the same. The ratio will be close to zero if the number of executes and parses Are almost equal. The ratio will be negative executes are lower but the parses are higher. % SQL with executions>1 value should also be high, a low value could mean that The database is not using shared SQL statements which in turn could mean that Bind variables are not being used. Top 5 Timed EventsTop 5 Timed Events Avg %Total ~~~~~~~~~~~~~~~~~~ wait Call Event Waits Time (s) (ms) Time Wait Class ------------------------------ ------------ ----------- ------ ------ ---------- control file parallel write 1,220 18 15 331.7 System I/O control file sequential read 6,508 6 1 110.9 System I/O CPU time 4 64.4 CGS wait for IPC msg 422,253 1 0 20.9 Other change tracking file synchrono 60 1 13 14.4 Other -------------------------------------------------------------
In an ideal database CPU and I/O should be the top wait events. If there are events like TX – row lock contention, Latch Free then that means there is contention in your database. If there is a high log file sync event then check why is it happening. A possible solution for this problem is to increase the size of the redo log buffer and to move your Logfiles to a storage system where Disk I/O capacity is high. Also, The db file sequential read(which means index reads) should be higher as compared to db file scattered read(which means full table scans). RAC Statistics If you are running on a RAC cluster, then the AWRRPT.SQL report will provide various RAC statistics including statistics on the number of RAC instances, as well as global cache and enqueue related performance statistics. Here is an example of the RAC statistics part of the report:RAC Statistics DB/Inst: A109/a1092 Snaps: 2009-2010 Begin End ----- ----- Number of Instances: 2 2 Global Cache Load Profile ~~~~~~~~~~~~~~~~~~~~~~~~~ Per Second Per Transaction --------------- --------------- Global Cache blocks received: 0.11 0.52 Global Cache blocks served: 0.14 0.68 GCS/GES messages received: 0.88 4.23 GCS/GES messages sent: 0.85 4.12 DBWR Fusion writes: 0.01 0.04 Estd Interconnect traffic (KB) 2.31 Global Cache Efficiency Percentages (Target local+remote 100%) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Buffer access - local cache %: 99.47 Buffer access - remote cache %: 0.53 Buffer access - disk %: 0.00 Global Cache and Enqueue Services - Workload Characteristics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Avg global enqueue get time (ms): 0.0 Avg global cache cr block receive time (ms): 0.2 Avg global cache current block receive time (ms): 0.3 Avg global cache cr block build time (ms): 0.0 Avg global cache cr block send time (ms): 0.0 Global cache log flushes for cr blocks served %: 1.8 Avg global cache cr block flush time (ms): 4.0 Avg global cache current block pin time (ms): 0.0 Avg global cache current block send time (ms): 0.1 Global cache log flushes for current blocks served %: 0.4 Avg global cache current block flush time (ms): 0.0 Global Cache and Enqueue Services - Messaging Statistics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Avg message sent queue time (ms): ######## Avg message sent queue time on ksxp (ms): 0.1 Avg message received queue time (ms): 4.6 Avg GCS message process time (ms): 0.0 Avg GES message process time (ms): 0.0 % of direct sent messages: 45.26 % of indirect sent messages: 31.59 % of flow controlled messages: 23.15 -------------------------------------------------------------
From the perspective of performance tuning the amount of block transfer across the interconnect should be low in comparison to localized access. This can be seen in the "Global Cache Efficiency Percentages" section. In the "Global Cache and Enqueue Services" the upper limit for Avg global cache cr block receive time should be 4 and Avg global cache current block receive time should be 8. Time Model Statistics Time related statistics presents the various operations which are consuming most of the database time.Time Model Statistics DB/Inst: A109/a1092 Snaps: 2009-2010 -> Total time in database user-calls (DB Time): 5.5s -> Statistics including the word "background" measure background process time, and so do not contribute to the DB time statistic -> Ordered by % or DB time desc, Statistic name Statistic Name Time (s) % of DB Time ------------------------------------------ ------------------ ------------ sql execute elapsed time 4.5 82.8 DB CPU 3.5 64.4 connection management call elapsed time 0.1 1.6 parse time elapsed 0.1 1.3 PL/SQL execution elapsed time 0.0 .9 hard parse elapsed time 0.0 .3 sequence load elapsed time 0.0 .1 repeated bind elapsed time 0.0 .0 DB time 5.5 N/A background elapsed time 33.0 N/A background cpu time 9.7 N/A -------------------------------------------------------------
If Hard parses or parsing time is very high then further investigation should be done to resolve the problem. Wait class and Wait Event Statistics Closely associated with the time model section of the report are the wait class and wait event statistics sections. Within Oracle, the duration of a large number of operations (e.g. Writing to disk or to the control file) is metered. These are known as wait events, because each of these operations requires the system to wait for the event to complete. Thus, the execution of some database operation (e.g. a SQL query) will have a number of wait events associated with it. We can try to determine which wait events are causing us problems by looking at the wait classes and the wait event reports generated from AWR. Wait classes define "buckets" that allow for summation of various wait times. Each wait event is assigned to one of these buckets (for example System I/O or User I/O). These buckets allow one to quickly determine which subsystem is likely suspect in performance problems (e.g. the network, or the cluster). Here is an example of the wait class report section:Wait Class DB/Inst: A109/a1092 Snaps: 2009-2010 -> s - second -> cs - centisecond - 100th of a second -> ms - millisecond - 1000th of a second -> us - microsecond - 1000000th of a second -> ordered by wait time desc, waits desc Avg %Time Total Wait wait Waits Wait Class Waits -outs Time (s) (ms) /txn -------------------- ---------------- ------ ---------------- ------- --------- System I/O 8,142 .0 25 3 10.9 Other 439,596 99.6 3 0 589.3 User I/O 112 .0 0 3 0.2 Cluster 443 .0 0 0 0.6 Concurrency 216 .0 0 0 0.3 Commit 16 .0 0 2 0.0 Network 3,526 .0 0 0 4.7 Application 13 .0 0 0 0.0 -------------------------------------------------------------
In this report the system I/O wait class has the largest number of waits (total of 25 seconds) and an average wait of 3 milliseconds. Wait events are normal occurrences, but if a particular sub-system is having a problem performing (e.g. the disk sub-system) this fact will appear in the form of one or more wait events with an excessive duration. The wait event report then provides some insight into the detailed wait events. Here is an example of the wait event report (we have eliminated some of the bulk of this report, because it can get quite long). Note that this section is sorted by wait time (listed in microseconds). Avg %Time Total Wait wait WaitsEvent Waits -outs Time (s) (ms) /txn ---------------------------- -------------- ------ ----------- ------- --------- control file parallel write 1,220 .0 18 15 1.6 control file sequential read 6,508 .0 6 1 8.7 CGS wait for IPC msg 422,253 100.0 1 0 566.0 change tracking file synchro 60 .0 1 13 0.1 db file parallel write 291 .0 0 1 0.4 db file sequential read 90 .0 0 4 0.1 reliable message 136 .0 0 1 0.2 log file parallel write 106 .0 0 2 0.1 lms flush message acks 1 .0 0 60 0.0 gc current block 2-way 200 .0 0 0 0.3 change tracking file synchro 59 .0 0 1 0.1
In this example our control file parallel write waits (which occurs during writes to the control file) are taking up 18 seconds total, with an average wait of 15 milliseconds per wait. Additionally we can see that we have 1.6 waits per transaction (or 15ms * 1.6 per transaction = 24ms). Operating System Statistics This part of the report provides some basic insight into OS performance, and OS configuration too. This report may vary depending on the OS platform that your database is running on. Here is an example from a Linux system:Statistic Total -------------------------------- -------------------- BUSY_TIME 128,749 IDLE_TIME 1,314,287 IOWAIT_TIME 18,394 NICE_TIME 54 SYS_TIME 31,633 USER_TIME 96,586 LOAD 0 RSRC_MGR_CPU_WAIT_TIME 0 PHYSICAL_MEMORY_BYTES 3,349,528 NUM_CPUS 4
In this example output, for example, we have 4 CPU's on the box. SQL In Need of Tuning Next in the report we find several different reports that present SQL statements that might be improved by tuning. There are a number of different reports that sort offending SQL statements by the following criteria: •Elapsed time •CPU time •Buffer gets •Physical reads •Executions •Parse calls •Sharable memory •Version count •Cluster wait time While these reports might not help tune specific application problems, they can help you find more systemic SQL problems that you might not find when tuning a specific application module. Here is an example of the Buffer gets report:Gets CPU Elapsed Buffer Gets Executions per Exec %Total Time (s) Time (s) SQL Id -------------- ------------ ------------ ------ -------- --------- ------------- 2,163 7 309.0 3.0 0.03 0.04 c7sn076yz7030 select smontabv.cnt, smontab.time_mp, smontab.scn, smontab.num_mappings, smon tab.tim_scn_map, smontab.orig_thread from smon_scn_time smontab, (sel ect max(scn) scnmax, count(*)+sum(NVL2(TIM_SCN_MAP,NUM_MAPPINGS, 0)) cnt from smon_scn_time where thread=0) smontabv where smon 1,442 721 2.0 2.0 0.05 0.05 6ssrk2dqj7jbx select job, nvl2(last_date, 1, 0) from sys.job$ where (((:1 <= next_date) and (n ext_date <= :2)) or ((last_date is null) and (next_date < :3))) and (field1 = :4 or (field1 = 0 and 'Y' = :5)) and (this_date is null) order by next_date, j ob 1,348 1 1,348.0 1.9 0.04 0.04 bv1djzzmk9bv6 Module: TOAD 9.0.0.160 Select table_name from DBA_TABLES where owner = 'CDOL2_01' order by 1 1,227 1 1,227.0 1.7 0.07 0.08 d92h3rjp0y217 begin prvt_hdm.auto_execute( :db_id, :inst_id, :end_snap ); end; 896 4 224.0 1.2 0.03 0.03 6hszmvz1wjhbt Module: TOAD 9.0.0.160 Select distinct Cons.constraint_name, cons.status, cons.table_name, cons.constra int_type ,cons.last_change from sys.user_constraints cons where 1=1 a nd cons.status='DISABLED'
In this report we find a SQL statement that seems to be churning through 309 buffers per execution. While the execution times are not terrible we might want to look closer into the SQL statement and try to see if we could tune it (in fact this is Oracle issued SQL that we would not tune anyway). Instance Activity Stats This section provides us with a number of various statistics (such as, how many DBWR Checkpoints occurred, or how many consistent gets occurred during the snapshot). Here is a partial example of the report:Statistic Total per Second per Trans -------------------------------- ------------------ -------------- ------------- consistent changes 9 0.0 0.0 consistent gets 70,445 19.5 94.4 consistent gets - examination 8,728 2.4 11.7 consistent gets direct 0 0.0 0.0 consistent gets from cache 70,445 19.5 94.4 cursor authentications 2 0.0 0.0 data blocks consistent reads - u 5 0.0 0.0 db block changes 1,809 0.5 2.4 db block gets 2,197 0.6 3.0 db block gets direct 0 0.0 0.0 db block gets from cache 2,033 0.6 2.7
Tablespace and Data File IO Stats The tablespace and data file IO stats report provides information on tablespace IO performance. From this report you can determine if the tablespace datafiles are suffering from sub-standard performance in terms of IO response from the disk sub-system. Here is a partial example of the tablespace report:Tablespace ------------------------------ Av Av Av Av Buffer Av Buf Reads Reads/s Rd(ms) Blks/Rd Writes Writes/s Waits Wt(ms) -------------- ------- ------ ------- ------------ -------- ---------- ------ SYSAUX 1 0 0.0 1.0 159 0 13 0.8 UNDOTBS2 1 0 10.0 1.0 98 0 0 0.0 SYSTEM 1 0 10.0 1.0 46 0 0 0.0 AUD 1 0 0.0 1.0 1 0 0 0.0 CDOL2_INDEX 1 0 10.0 1.0 1 0 0 0.0 CDOL_DATA 1 0 10.0 1.0 1 0 0 0.0 DBA_DEF 1 0 10.0 1.0 1 0 0 0.0 UNDOTBS1 1 0 10.0 1.0 1 0 0 0.0 USERS 1 0 10.0 1.0 1 0 0 0.0 USER_DEF 1 0 10.0 1.0 1 0 0 0.0
If the tablespace IO report seems to indicate a tablespace has IO problems, we can then use the file IO stat report allows us to drill into the datafiles of the tablespace in question and determine what the problem might be. Here is an example of the File IO stat report:Tablespace Filename ------------------------ ---------------------------------------------------- Av Av Av Av Buffer Av Buf Reads Reads/s Rd(ms) Blks/Rd Writes Writes/s Waits Wt(ms) -------------- ------- ------ ------- ------------ -------- ---------- ------ AUD +ASM01/a109/datafile/aud.296.604081931 1 0 0.0 1.0 1 0 0 0.0 CDOL2_INDEX +ASM01/a109/datafile/cdol2_index_001.dbf 1 0 10.0 1.0 1 0 0 0.0 CDOL_DATA +ASM01/a109/datafile/cdol_data_001.dbf 1 0 10.0 1.0 1 0 0 0.0 DBA_DEF +ASM01/a109/datafile/dba_def.294.604081931 1 0 10.0 1.0 1 0 0 0.0 SYSAUX +ASM01/a109/datafile/sysaux.299.604081927 1 0 0.0 1.0 159 0 13 0.8 SYSTEM +ASM01/a109/datafile/system.301.604081919 1 0 10.0 1.0 46 0 0 0.0 UNDOTBS1 +ASM01/a109/datafile/undotbs1.300.604081925 1 0 10.0 1.0 1 0 0 0.0 UNDOTBS2 +ASM01/a109/datafile/undotbs2.292.604081931 1 0 10.0 1.0 98 0 0 0.0 USERS +ASM01/a109/datafile/users.303.604081933 1 0 10.0 1.0 1 0 0 0.0 USER_DEF +ASM01/a109/datafile/user_def.291.604081933 1 0 10.0 1.0 1 0 0 0.0 -------------------------------------------------------------
Buffer Pool Statistics The buffer pool statistics report follows. It provides a summary of the buffer pool configuration and usage statistics as seen in this example:Free Writ Buffer Number of Pool Buffer Physical Physical Buff Comp Busy P Buffers Hit% Gets Reads Writes Wait Wait Waits --- ---------- ---- -------------- ------------ ----------- ---- ---- ---------- D 64,548 100 72,465 0 355 0 0 13 -------------------------------------------------------------
In this case, we have a database where all the buffer pool requests came out of the buffer pool and no physical reads were required. We also see a few (probably very insignificant in our case) buffer busy waits. Instance Recovery Stats The instance recovery stats report provides information related to instance recovery. By analyzing this report, you can determine roughly how long your database would have required to perform crash recovery during the reporting period. Here is an example of this report:-> B: Begin snapshot, E: End snapshot Targt Estd Log File Log Ckpt Log Ckpt MTTR MTTR Recovery Actual Target Size Timeout Interval (s) (s) Estd IOs Redo Blks Redo Blks Redo Blks Redo Blks Redo Blks - ----- ----- ---------- --------- --------- ---------- --------- ------------ B 0 19 196 575 183 92160 183 N/A E 0 19 186 258 96 92160 96 N/A -------------------------------------------------------------
Buffer Pool Advisory The buffer pool advisory report answers the question, how big should you make your database buffer cache. It provides an extrapolation of the benefit or detriment that would result if you added or removed memory from the database buffer cache. These estimates are based on the current size of the buffer cache and the number of logical and physical IO's encountered during the reporting point. This report can be very helpful in "rightsizing" your buffer cache. Here is an example of the output of this report:Est Phys Size for Size Buffers for Read Estimated P Est (M) Factor Estimate Factor Physical Reads --- -------- ------ ---------------- ------ ------------------ D 48 .1 5,868 4.9 803,496 D 96 .2 11,736 4.0 669,078 D 144 .3 17,604 3.3 550,831 D 192 .4 23,472 2.8 462,645 D 240 .5 29,340 2.3 379,106 D 288 .5 35,208 1.8 305,342 D 336 .6 41,076 1.4 238,729 D 384 .7 46,944 1.2 200,012 D 432 .8 52,812 1.1 183,694 D 480 .9 58,680 1.0 172,961 D 528 1.0 64,548 1.0 165,649 D 576 1.1 70,416 1.0 161,771 D 624 1.2 76,284 1.0 159,728 D 672 1.3 82,152 1.0 158,502 D 720 1.4 88,020 1.0 157,723 D 768 1.5 93,888 0.9 157,124 D 816 1.5 99,756 0.9 156,874 D 864 1.6 105,624 0.9 156,525 D 912 1.7 111,492 0.9 156,393 D 960 1.8 117,360 0.9 155,388 -------------------------------------------------------------
In this example we currently have 528GB allocated to the SGA (represented by the size factor column with a value of 1.0. It appears that if we were to reduce the memory allocated to the SGA to half of the size of the current SGA (freeing the memory to the OS for other processes) we would incur an increase of about 1.8 times the number of physical IO's in the process. PGA Reports The PGA reports provide some insight into the health of the PGA. The PGA Aggr Target Stats report provides information on the configuration of the PGA Aggregate Target parameter during the reporting period. The PGA Aggregate Target Histogram report provides information on the size of various operations (e.g. sorts). It will indicate if PGA sort operations occurred completely in memory, or if some of those operations were written out to disk. Finally the PGA Memory Advisor, much like the buffer pool advisory report, provides some insight into how to properly size your PGA via the PGA_AGGREGATE_TARGET database parameter. The PGA Memory Advisor report is shown here:Estd Extra Estd PGA Estd PGA PGA Target Size W/A MB W/A MB Read/ Cache Overalloc Est (MB) Factr Processed Written to Disk Hit % Count ---------- ------- ---------------- ---------------- -------- ---------- 44 0.1 289,899.2 7,844.9 97.0 1,124 88 0.3 289,899.2 7,576.9 97.0 1,073 176 0.5 289,899.2 3.3 100.0 0 263 0.8 289,899.2 3.3 100.0 0 351 1.0 289,899.2 3.3 100.0 0 421 1.2 289,899.2 0.0 100.0 0 491 1.4 289,899.2 0.0 100.0 0 562 1.6 289,899.2 0.0 100.0 0 632 1.8 289,899.2 0.0 100.0 0 702 2.0 289,899.2 0.0 100.0 0 1,053 3.0 289,899.2 0.0 100.0 0 1,404 4.0 289,899.2 0.0 100.0 0 2,106 6.0 289,899.2 0.0 100.0 0 2,808 8.0 289,899.2 0.0 100.0 0 -------------------------------------------------------------
Shared Pool Advisory The shared pool advisory report provides assistance in right sizing the Oracle shared pool. Much like the PGA Memory Advisor or the Buffer Pool advisory report, it provides some insight into what would happen should you add or remove memory from the shared pool. This can help you reclaim much needed memory if you have over allocated the shared pool, and can significantly improve performance if you have not allocated enough memory to the shared pool. Here is an example of the shared pool advisory report:Est LC Est LC Est LC Est LC Shared SP Est LC Time Time Load Load Est LC Pool Size Size Est LC Saved Saved Time Time Mem Size(M) Factr (M) Mem Obj (s) Factr (s) Factr Obj Hits ---------- ----- -------- ------------ ------- ------ ------- ------ ----------- 192 .4 54 3,044 ####### .8 ####### 382.1 22,444,274 240 .5 92 5,495 ####### .9 ####### 223.7 22,502,102 288 .6 139 8,122 ####### .9 53,711 102.5 22,541,782 336 .7 186 12,988 ####### 1.0 17,597 33.6 22,562,084 384 .8 233 17,422 ####### 1.0 7,368 14.1 22,569,402 432 .9 280 23,906 ####### 1.0 3,553 6.8 22,571,902 480 1.0 327 28,605 ####### 1.0 524 1.0 22,573,396 528 1.1 374 35,282 ####### 1.0 1 .0 22,574,164 576 1.2 421 40,835 ####### 1.0 1 .0 22,574,675 624 1.3 468 46,682 ####### 1.0 1 .0 22,575,055 672 1.4 515 52,252 ####### 1.0 1 .0 22,575,256 720 1.5 562 58,181 ####### 1.0 1 .0 22,575,422 768 1.6 609 64,380 ####### 1.0 1 .0 22,575,545 816 1.7 656 69,832 ####### 1.0 1 .0 22,575,620 864 1.8 703 75,168 ####### 1.0 1 .0 22,575,668 912 1.9 750 78,993 ####### 1.0 1 .0 22,575,695 960 2.0 797 82,209 ####### 1.0 1 .0 22,575,719 -------------------------------------------------------------
SGA Target Advisory The SGA target advisory report is somewhat of a summation of all the advisory reports previously presented in the AWR report. It helps you determine the impact of changing the settings of the SGA target size in terms of overall database performance. The report uses a value called DB Time as a measure of the increase or decrease in performance relative to the memory change made. Also the report will summarize an estimate of physical reads associated with the listed setting for the SGA. Here is an example of the SGA target advisory report:SGA Target SGA Size Est DB Est Physical Size (M) Factor Time (s) Reads ---------- ---------- ------------ ---------------- 528 0.5 25,595 769,539 792 0.8 20,053 443,095 1,056 1.0 18,443 165,649 1,320 1.3 18,354 150,476 1,584 1.5 18,345 148,819 1,848 1.8 18,345 148,819 2,112 2.0 18,345 148,819
In this example, our SGA Target size is currently set at 1056MB. We can see from this report that if we increased the SGA target size to 2112MB, we would see almost no performance improvement (about a 98 second improvement overall). In this case, we may determine that adding so much memory to the database is not cost effective, and that the memory can be better used elsewhere. Memory Advisory Memory advisory reports for the streams pool and the java pool also appear in the report (assuming you are using the streams pool). These reports take on the same general format as the other memory advisor reports. Buffer Wait Statistics The buffer wait statistics report helps you drill down on specific buffer wait events, and where the waits are occurring. In the following report we find that the 13 buffer busy waits we saw in the buffer pool statistics report earlier are attributed to data block waits. We might then want to pursue tuning remedies to these waits if the waits are significant enough. Here is an example of the buffer wait statistics report:Class Waits Total Wait Time (s) Avg Time (ms) ------------------ ----------- ------------------- -------------- data block 13 0 1
Enqueue Activity The Enqueue activity report provides information on enqueues (higher level Oracle locking) that occur. As with other reports, if you see high levels of wait times in these reports, you might dig further into the nature of the enqueue and determine the cause of the delays. Here is an example of this report section: Enqueue Type (Request Reason)------------------------------------------------------------------------------ Requests Succ Gets Failed Gets Waits Wt Time (s) Av Wt Time(ms) ------------ ------------ ----------- ----------- ------------ -------------- PS-PX Process Reservation 386 358 28 116 0 .43 US-Undo Segment 276 276 0 228 0 .18 TT-Tablespace 90 90 0 42 0 .71 WF-AWR Flush 12 12 0 7 0 1.43 MW-MWIN Schedule 2 2 0 2 0 5.00 TA-Instance Undo 12 12 0 12 0 .00 UL-User-defined 7 7 0 7 0 .00 CF-Controlfile Transaction 5,737 5,737 0 5 0 .00
Undo Segment Summary The undo segment summary report provides basic information on the performance of undo tablespaces. Latch Activity The latch activity report provides information on Oracle's low level locking mechanism called a latch. From this report you can determine if Oracle is suffering from latching problems, and if so, which latches are causing the greates amount of contention on the system. Here is a partial example of the latch activity report (it is quite long):Pct Avg Wait Pct Get Get Slps Time NoWait NoWait Latch Name Requests Miss /Miss (s) Requests Miss ------------------------ -------------- ------ ------ ------ ------------ ------ ASM allocation 122 0.0 N/A 0 0 N/A ASM map headers 60 0.0 N/A 0 0 N/A ASM map load waiting lis 11 0.0 N/A 0 0 N/A ASM map operation freeli 30 0.0 N/A 0 0 N/A ASM map operation hash t 45,056 0.0 N/A 0 0 N/A ASM network background l 1,653 0.0 N/A 0 0 N/A AWR Alerted Metric Eleme 14,330 0.0 N/A 0 0 N/A Consistent RBA 107 0.0 N/A 0 0 N/A FAL request queue 75 0.0 N/A 0 0 N/A FAL subheap alocation 75 0.0 N/A 0 0 N/A FIB s.o chain latch 14 0.0 N/A 0 0 N/A FOB s.o list latch 93 0.0 N/A 0 0 N/A JS broadcast add buf lat 826 0.0 N/A 0 0 N/A JS broadcast drop buf la 826 0.0 N/A 0 0 N/A
In this example our database does not seem to be experiencing any major latch problems, as the wait times on the latches are 0, and our get miss pct (Pct Get Miss) is 0 also. There is also a latch sleep breakdown report which provides some additional detail if a latch is being constantly moved into the sleep cycle, which can cause additional performance issues. The latch miss sources report provides a list of latches that encountered sleep conditions. This report can be of further assistance when trying to analyze which latches are causing problems with your database. Segments by Logical Reads and Segments by Physical Reads The segments by logical reads and segments by physical reads reports provide information on the database segments (tables, indexes) that are receiving the largest number of logical or physical reads. These reports can help you find objects that are "hot" objects in the database. You may want to review the objects and determine why they are hot, and if there are any tuning opportunities available on those objects (e.g. partitioning), or on SQL accessing those objects. For example, if an object is showing up on the physical reads report, it may be that an index is needed on that object. Here is an example of the segments by logical reads report:Segments by Logical Reads DB/Inst: A109/a1092 Snaps: 2009-2010 -> Total Logical Reads: 72,642 -> Captured Segments account for 96.1% of Total Tablespace Subobject Obj. Logical Owner Name Object Name Name Type Reads %Total ---------- ---------- -------------------- ---------- ----- ------------ ------- SYS SYSAUX SYS_IOT_TOP_8813 INDEX 52,192 71.85 SYS SYSTEM SMON_SCN_TIME TABLE 4,704 6.48 SYS SYSTEM I_JOB_NEXT INDEX 2,432 3.35 SYS SYSTEM OBJ$ TABLE 1,344 1.85 SYS SYSTEM TAB$ TABLE 1,008 1.39 -------------------------------------------------------------
Additional Reports Several segment related reports appear providing information on: •Segments with ITL waits •Segments with Row lock waits •Segments with buffer busy waits •Segments with global cache buffer waits •Segments with CR Blocks received •Segments with current blocks received These reports help provide more detailed information on specific segments that might be experiencing performance problems. The dictionary cache and library cache statistics reports provide performance information on the various areas in the data dictionary cache and the library cache. The process memory summary, SGA memory summary, and the SGA breakdown difference reports provide summary information on how memory allocated to the database is allocated amongst the various components. Other memory summary reports may occur if you have certain optional components installed (such as streams). The database parameter summary report provides a summary of the setting of all the database parameters during the snapshot report. If the database parameters changed during the period of the report, then the old and new parameters will appear on the report. Related: - 1) How to Tune Shared Pool Cache ? Reference: ToadWorld
Search This Blog
Saturday, May 5, 2012
How To Read An AWR Report ?
Subscribe to:
Post Comments (Atom)
really good and informative
ReplyDeleteThis is best article in AWR I have ever read...
ReplyDeleteThanks for sharing.
Cool!!! Thanks for this article :)))))
ReplyDeleteThis is excellent Article to understand and how analize AWR report for beginners.
ReplyDeleteThanks for sharing this valuble info to others.
Regards
Jagadeesh
great material for beginners, thank you
ReplyDeleteNice post .....very informative........!!!
ReplyDelete