| Oracle9i Database Performance Tuning Guide and Reference Release 2 (9.2) Part Number A96533-02 |
|
|
View PDF |
| Oracle9i Database Performance Tuning Guide and Reference Release 2 (9.2) Part Number A96533-02 |
|
|
View PDF |
After the initial configuration of a database, tuning an instance is important to eliminate any performance bottlenecks.
This chapter contains the following sections:
Performance tuning requires a different, although related, method to the initial configuration of a system. Configuring a system involves allocating resources in an ordered manner so that the initial system configuration is functional.
Tuning is driven by identifying the most significant bottleneck and making the appropriate changes to reduce or eliminate the effect of that bottleneck. Usually, tuning is performed reactively, either while the system is preproduction or after it is live.
|
Note: Before using this performance tuning reference, make sure you have read Oracle9i Database Performance Planning. Oracle Corporation has designed a new performance methodology, based on years of Oracle designing and performance experience. This brief book explains clear and simple activities that can dramatically improve system performance. It discusses the following topics:
|
The most effective way to tune is to have an established performance baseline that can be used for comparison if a performance issue arises. Most DBAs know their system well and can easily identify peak usage periods. For example, the peak periods could be between 10.00am and 12.00pm and also between 1.30pm and 3.00pm. This could include a batch window of 12.00am midnight to 6am.
It is important to identify these high-load times at the site and install a monitoring tool that gathers performance data for those times. Optimally, data gathering should be configured from when the application is in its initial trial phase during the QA cycle. Otherwise, this should be configured when the system is first in production.
| See Also:
Chapter 20, "Oracle Tools to Gather Database Statistics" for detailed information on Oracle instance performance tuning tools |
Ideally, baseline data gathered should include the following:
A common pitfall in performance tuning is to mistake the symptoms of a problem for the actual problem itself. It is important to recognize that many performance statistics indicate the symptoms, and that identifying the symptom is not sufficient data to implement a remedy. For example:
Generally, this is caused by poorly-configured disks. However, it could also be caused by a significant amount of unnecessary physical I/O on those disks issued by poorly-tuned SQL.
Rarely is latch contention tunable by reconfiguring the instance. Rather, latch contention usually is resolved through application changes.
Excessive CPU usage usually means that there is little idle CPU on the system. This could be caused by an inadequately-sized system, by untuned SQL statements, or by inefficient application programs.
There are two distinct types of tuning: proactive monitoring and bottleneck elimination.
Proactive monitoring usually occurs on a regularly scheduled interval, where a number of performance statistics are examined to identify whether the system behavior and resource usage has changed. Proactive monitoring also can be called proactive tuning.
Usually, monitoring does not result in configuration changes to the system, unless the monitoring exposes a serious problem that is developing. In some situations, experienced performance engineers can identify potential problems through statistics alone, although accompanying performance degradation is usual.
'Tweaking' a system when there is no apparent performance degradation as a proactive action can be a dangerous activity, resulting in unnecessary performance drops. Tweaking a system should be considered reactive tuning, and the steps for reactive tuning should be followed.
Monitoring is usually part of a larger capacity planning exercise, where resource consumption is examined to see the changes in the way the application is being used and the way the application is using the database and host resources.
Tuning usually implies fixing a performance problem. However, tuning should be part of the lifecycle of an application, through the analysis, design, coding, production, and maintenance stages. Many times, the tuning phase is left until the system is in production. At this time, tuning becomes a reactive fire-fighting exercise, where the most important bottleneck is identified and fixed.
Usually, the purpose for tuning is to reduce resource consumption or to reduce the elapsed time for an operation to complete. Either way, the goal is to improve the effective use of a particular resource. In general, performance problems are caused by the over-use of a particular resource. That resource is the bottleneck in the system. There are a number of distinct phases in identifying the bottleneck and the potential fixes. These are discussed in the sections that follow.
Remember that the different forms of contention are symptoms that can be fixed by making changes in the following places:
Often, the most effective way of resolving a bottleneck is to change the application.
These are the main steps in the Oracle Performance Method:
| See Also:
Oracle9i Database Performance Planning for a list of common errors and for a theoretical description of this performance method |
The remainder of this chapter covers the steps of the Oracle performance method in detail.
It is vital to develop a good understanding of the purpose of the tuning exercise and the nature of the problem before attempting to implement a solution. Without this understanding, it is virtually impossible to implement effective changes. The data gathered during this stage helps determine the next step to take and what evidence to examine.
Gather the following data:
What is the measure of acceptable performance? How many transactions an hour, or seconds, response time will meet the required performance level?
What is affected by the slowdown? For example, is the whole instance slow? Is it a particular application, program, specific operation, or a single user?
Is the problem only evident during peak hours? Does performance deteriorate over the course of the day? Was the slowdown gradual (over the space of months or weeks) or sudden?
This helps identify the extent of the problem and also acts as a measure for comparison when deciding whether changes implemented to fix the problem have actually made an improvement. Find a consistently reproducible measure of the response time or job run time. How much worse are the timings than when the program was running well?
Identify what has changed since performance was acceptable. This may narrow the potential cause quickly. For example, has the operating system software, hardware, application software, or Oracle release been upgraded? Has more data been loaded into the system, or has the data volume or user population grown?
At the end of this phase, you should have a good understanding of the symptoms. If the symptoms can be identified as local to a program or set of programs, then the problem is handled in a different manner than instance-wide performance issues.
| See Also:
Chapter 6, "Optimizing SQL Statements" for information on solving performance problems specific to an application or user |
Look at the load on the database server, as well as the database instance. Consider the operating system, the I/O subsystem, and network statistics, because examining these areas helps determine what might be worth further investigation. In multitier systems, also examine the application server middle-tier hosts.
Examining the host hardware often gives a strong indication of the bottleneck in the system. This determines which Oracle performance data could be useful for cross-reference and further diagnosis.
Data to examine includes the following:
If there is a significant amount of idle CPU, then there could be an I/O, application, or database bottleneck. Note that wait I/O should be considered as idle CPU.
If there is high CPU usage, then determine whether the CPU is being used effectively. Is the majority of CPU usage attributable to a small number of high-CPU using programs, or is the CPU consumed by an evenly distributed workload?
If the CPU is used by a small number of high-usage programs, then look at the programs to determine the cause.
If the programs are not Oracle programs, then identify whether they are legitimately requiring that amount of CPU. If so, then can their execution can be delayed to off-peak hours?
If a small number of Oracle processes consumes most of the CPU resources, then use SQL_TRACE and TKPROF to identify the SQL or PL/SQL statements to see if a particular query or PL/SQL program unit can be tuned. For example, a SELECT statement could be CPU-intensive if its execution involves many reads of data in cache (logical reads) that could be avoided with better SQL optimization.
Oracle CPU statistics are available in three V$ views:
V$SYSSTAT shows Oracle CPU usage for all sessions. The statistic "CPU used by this session" shows the aggregate CPU used by all sessions.V$SESSTAT shows Oracle CPU usage for each session. Use this view to determine which particular session is using the most CPU.V$RSRC_CONSUMER_GROUP shows CPU utilization statistics for each consumer group when the Oracle Database Resource Manager is running.It is important to recognize that CPU time and real time are distinct. With eight CPUs, for any given minute in real time, there are eight minutes of CPU time available. On NT and UNIX, this can be either user time or system time (privileged mode on NT). Thus, average CPU time utilized by all processes (threads) on the system could be greater than one minute for every one minute real time interval.
At any given moment, you know how much time Oracle has used on the system. So, if eight minutes are available and Oracle uses four minutes of that time, then you know that 50% of all CPU time is used by Oracle. If your process is not consuming that time, then some other process is. Identify the processes that are using CPU time, figure out why, and then attempt to tune them.
If the CPU usage is evenly distributed over many Oracle server processes, then examine the Statspack report for other evidence.
An overly active I/O system can be evidenced by disk queue lengths greater than two, or disk service times that are over 20-30ms. If the I/O system is overly active, then check for potential hot spots that could benefit from distributing the I/O across more disks. Also identify whether the load can be reduced by lowering the resource requirements of the programs using those resources.
Use operating system monitoring tools to determine what processes are running on the system as a whole and to monitor disk access to all files. Remember that disks holding datafiles and redo log files can also hold files that are not related to Oracle. Reduce any heavy access to disks that contain database files. Access to non-Oracle files can be monitored only through operating system facilities, rather than through the V$ views.
Tools, such as sar -d (or iostat) on many UNIX systems and Performance Monitor on Windows 2000 systems, examine I/O statistics for the entire system.
Check the Oracle wait event data in V$SYSTEM_EVENT to see whether the top wait events are I/O related. I/O related events include db file sequential read, db file scattered read, db file single write, and db file parallel write. These are all events corresponding to I/Os performed against the data file headers, control files, or data files. If any of these wait events correspond to high average time, then investigate the I/O contention.
Cross reference the host I/O system data with the I/O sections in the Statspack report to identify hot datafiles and tablespaces. Also compare the I/O times reported by the operating system with the times reported by Oracle to see if they are consistent.
Before investigating whether the I/O system should be reconfigured, determine if the load on the I/O system can be reduced. To reduce Oracle I/O load, look at SQL statements that perform many physical reads by querying the V$SQLAREA view or by reviewing the 'SQL ordered by physical reads' section of the Statspack report. Examine these statements to see how they can be tuned to reduce the number of I/Os.
If there are Oracle-related I/O problems caused by SQL statements, then tune them. If the Oracle server is not consuming the available I/O resources, then identify the process that is using up the I/O. Determine why the process is using up the I/O, and then tune this process.
See Also:
|
Using operating system utilities, look at the network round-trip ping time and the number of collisions. If the network is causing large delays in response time, then investigate possible causes.
Network load can be reduced by scheduling large data transfers to off-peak times, or by coding applications to batch requests to remote hosts, rather than accessing remote hosts once (or more) for one request.
| See Also:
Oracle9i Database Performance Planning for a description of important operating system statistics |
Oracle statistics should be examined and cross-referenced with-operating system statistics to ensure a consistent diagnosis of the problem. operating-system statistics can indicate a good place to begin tuning. However, if the goal is to tune the Oracle instance, then look at the Oracle statistics to identify the resource bottleneck from Oracle's perspective before implementing corrective action.
Oracle9i Release 2 (9.2) provides the initialization parameter STATISTICS_LEVEL, which controls all major statistics collections or advisories in the database. This parameter sets the statistics collection level for the database.
Depending on the setting of STATISTICS_LEVEL, certain advisories and statistics are collected, as follows:
BASIC: No advisories or statistics are collected.
TYPICAL: The following advisories or statistics are collected:
ALL: All of the preceding advisories or statistics are collected, plus the following:
The default level is TYPICAL. STATISTICS_LEVEL is a dynamic parameter and can be altered at the system or the session level.
When modified by ALTER SYSTEM, all advisories or statistics in the preceding list are dynamically turned on or off, depending on the new value of STATISTICS_LEVEL.
When modified by ALTER SESSION, only the following advisories or statistics are turned on or off in the local session only. Their system-wide state is not changed.
This view lists the status of the statistics or advisories controlled by STATISTICS_LEVEL.
| See Also:
"V$STATISTICS_LEVEL" for column details of this view |
The following sections discuss the common Oracle data sources used while tuning. The sources can be divided into two types of statistics: wait events and system statistics.
Wait events are statistics that are incremented by a server process/thread to indicate that it had to wait for an event to complete before being able to continue processing. Wait event data reveals various symptoms of problems that might be impacting performance, such as latch contention, buffer contention, and I/O contention. Remember that these are only symptoms of problems--not the actual causes.
A server process can wait for the following:
Wait event statistics include the number of times an event was waited for and the time waited for the event to complete. The views V$SESSION_WAIT, V$SESSION_EVENT, and V$SYSTEM_EVENT can be queried for wait event statistics. If the configuration parameter TIMED_STATISTICS is set to true, then you can also see how long each resource was waited for. To minimize user response time, reduce the time spent by server processes waiting for event completion. Not all wait events have the same wait time. Therefore, it is more important to examine events with the most total time waited rather than wait events with a high number of occurrences. Usually, it is best to set the dynamic parameter TIMED_STATISTICS to true at least while monitoring performance.
| See Also:
"Setting the Level of Statistics Collection" for information about |
Investigate wait events and related timing data when performing reactive performance tuning. The events with the most time listed against them are often strong indications of the performance bottleneck. For example, by looking at V$SYSTEM_EVENT, you might notice lots of buffer busy waits. It might be that many processes are inserting into the same block and must wait for each other before they can insert. The solution could be to introduce freelists for the object in question.
| See Also:
See Wait Events for a description of the differences between the views |
System statistics are typically used in conjunction with wait event data to find further evidence of the cause of a performance problem.
For example, if V$SYSTEM_EVENT indicates that the largest wait event (in terms of wait time) is the event buffer busy waits, then look at the specific buffer wait statistics available in the view V$WAITSTAT to see which block type has the highest wait count and the highest wait time. After the block type has been identified, also look at V$SESSION_WAIT real-time while the problem is occurring to identify the contended-for object(s) using the file number and block number indicated. The combination of this data indicates the appropriate corrective action.
Statistics are available in many V$ views. Some common views include the following:
This contains overall statistics for many different parts of Oracle, including rollback, logical and physical I/O, and parse data. Data from V$SYSSTAT is used to compute ratios, such as the buffer cache hit ratio.
This contains detailed file I/O statistics for each file, including the number of I/Os for each file and the average read time.
This contains detailed rollback and undo segment statistics for each segment.
This contains detailed enqueue statistics for each enqueue, including the number of times an enqueue was requested and the number of times an enqueue was waited for, and the wait time.
This contains detailed latch usage statistics for each latch, including the number of times each latch was requested and the number of times the latch was waited for.
| See Also:
Chapter 24, "Dynamic Performance Views for Tuning" for detailed descriptions of the |
With Oracle9i Release 2 (9.2) and higher, you can gather segment-level statistics to help you spot performance problems associated with individual segments. Collecting and viewing segment-level statistics is a good way to effectively identify the hot table or index in an instance.
After viewing wait events or system statistics to identify the performance problem, you can use segment-level statistics to find specific tables or indexes that are causing the problem. Consider, for example, that V$SYSTEM_EVENT indicates that buffer busy waits cause a fair amount of wait time. You can select from V$SEGMENT_STATISTICS the top segments that cause the buffer busy waits. Then you can focus your effort on eliminating the problem in those segments.
You can query segment-level statistics through the following dynamic performance views:
V$SEGSTAT_NAME This view lists the segment statistics being collected, as well as the properties of each statistic (for instance, if it is a sampled statistic).V$SEGSTAT This is a highly efficient, real-time monitoring view that shows the statistic value, statistic name, and other basic information.V$SEGMENT_STATISTICS This is a user-friendly view of statistic values. In addition to all the columns of V$SEGSTAT, it has information about such things as the segment owner and table space name. It makes the statistics easy to understand, but it is more costly.
| See Also:
Chapter 24, "Dynamic Performance Views for Tuning" for detailed column information, starting |
Often at the end of a tuning exercise, it is possible to identify two or three changes that could potentially alleviate the problem. To identify which change provides the most benefit, it is recommended that only one change be implemented at a time. The effect of the change should measured against the baseline data measurements found in the problem definition phase.
Typically, most sites with dire performance problems implement a number of overlapping changes at once, and thus cannot identify which changes provided any benefit. Although this is not immediately an issue, this becomes a significant hindrance if similar problems subsequently appear, because it is not possible to know which of the changes provided the most benefit and which efforts to prioritize.
If it is not possible to implement changes separately, then try to measure the effects of dissimilar changes. For example, measure the effect of making an initialization change to optimize redo generation separately from the effect of creating a new index to improve the performance of a modified query. It is impossible to measure the benefit of performing an operating system upgrade if SQL is tuned, the operating system disk layout is changed, and the initialization parameters are also changed at the same time.
Performance tuning is an iterative process. It is unlikely to find a 'silver bullet' that solves an instance-wide performance problem. In most cases, excellent performance requires iteration through the performance tuning phases, because solving one bottleneck often uncovers another (sometimes worse) problem.
Knowing when to stop tuning is also important. The best measure of performance is user perception, rather than how close the statistic is to an ideal value.
Gather statistics that cover the time when the instance had the performance problem. If you previously captured baseline data for comparison, then you can compare the current data to the data from the baseline that most represents the problem workload.
When comparing two reports, ensure that the two reports are from times where the system was running comparable workloads.
Usually, wait events are the first data examined. However, if you have a baseline report, then check to see if the load has changed. Regardless of whether you have a baseline, it is useful to see whether the resource usage rates are high.
Load-related statistics to examine include redo size, session logical reads, db block changes, physical reads, physical writes, parse count (total), parse count (hard), and user calls. This data is queried from V$SYSSTAT. It is best to normalize this data over seconds and over transactions.
In the Statspack report, look at the Load Profile section. The data has been normalized over transactions and over seconds.
The load profile statistics over seconds show the changes in throughput (that is, whether the instance is performing more work each second). The statistics over transactions identify changes in the application characteristics by comparing these to the corresponding statistics from the baseline report.
Examine the statistics normalized over seconds to identify whether the rates of activity are very high. It is difficult to make blanket recommendations on high values, because the thresholds are different on each site and are contingent on the application characteristics, the number and speed of CPUs, the operating system, the I/O system, and the Oracle release.
The following are some generalized examples (acceptable values vary at each site):
Whenever an Oracle process waits for something, it records the wait using one of a set of predefined wait events. (See V$EVENT_NAME for a list of all wait events.) Some of these events are termed idle events, because the process is idle, waiting for work to perform. Non-idle events indicate nonproductive time spent waiting for a resource or action to complete.
|
Note: Not all symptoms can be evidenced by wait events. See "Additional Statistics" for the statistics that can be checked. |
The most effective way to use wait event data is to order the events by the wait time. This is only possible if TIMED_STATISTICS is set to true. Otherwise, the wait events can only be ranked by the number of times waited, which is often not the ordering that best represents the problem.
| See Also:
"Setting the Level of Statistics Collection" for information about |
To get an indication of where time is spent, follow these steps:
V$SYSTEM_EVENT. The events of interest should be ranked by wait time.
Identify the wait events that have the most significant percentage of wait time. To determine the percentage of wait time, add the total wait time for all wait events, excluding idle events (such as Null event, SQL*Net message from client, SQL*Net message to client, SQL*Net more data). Calculate the relative percentage of the five most prominent events by dividing each event's wait time by the total time waited for all events.
.| See Also:
"Idle Wait Events" for the complete list of idle events |
Alternatively, look at the Top 5 Wait Events section on the front page of the Statspack report; this section automatically orders the wait events (omitting idle events), and calculates the relative percentage:
Top 5 Wait Events ~~~~~~~~~~~~~~~~~ Wait % Total Event Waits Time (cs) Wt Time ------------------------------------------ ------------ ------------ ------- latch free 217,224 65,056 63.55 db file sequential read 39,836 31,844 31.11 db file scattered read 3,679 2,846 2.78 SQL*Net message from dblink 1,186 870 .85 log file sync 830 775 .76
In the previous example, the highest ranking wait event is the latch free event. In some situations, there might be a few events with similar percentages. This can provide extra evidence if all the events are related to the same type of resource request (for example, all I/O related events).
Avg Total Wait wait Waits Event Waits Timeouts Time (s) (ms) /txn -------------------------- ------------ ---------- ----------- ----- ------ latch free 5,560,989 2,705,969 26,117 5 827.7 db file sequential read 137,027 0 2,129 16 20.4 SQL*Net break/reset to cli 1,566 0 1,707 1091 0.2
V$SYSSTAT, operating system statistics, and so on). Perform cross-checks with other data to confirm or refute the developing theory.Table 22-1 links wait events to possible causes and gives an overview of the Oracle data that could be most useful to review next.
See Also:
|
There are a number of statistics that can indicate performance problems that do not have corresponding wait events.
The V$SYSSTAT statistic redo log space requests indicates how many times a server process had to wait for space in the online redo log, not for space in the redo log buffer. A significant value for this statistic and the wait events should be used as an indication that checkpoints, DBWR, or archiver activity should be tuned, not LGWR. Increasing the size of log buffer does not help.
Your system might spend excessive time rolling back changes to blocks in order to maintain a consistent view. Consider the following scenarios:
V$SYSSTAT statistics to determine whether this is happening:
V$SYSSTAT statistics should be close to 1:
ratio = transaction tables consistent reads undo records applied / transaction tables consistent read rollbacks
A solution is to create more, smaller rollback segments, or to use automatic undo management.
WAITS to the number of GETS in V$ROLLSTAT; the proportion of WAITS to GETS should be small.V$WAITSTAT to see whether there are many WAITS for buffers of CLASS 'undo header'.A solution is to create more rollback segments.
You can detect migrated or chained rows by checking the number of table fetch continued row statistic in V$SYSSTAT. A small number of chained rows (less than 1%) is unlikely to impact system performance. However, a large percentage of chained rows can affect performance.
Chaining on rows larger than the block size is inevitable. You might want to consider using tablespace with larger block size for such data.
However, for smaller rows, you can avoid chaining by using sensible space parameters and good application design. For example, do not insert a row with key values filled in and nulls in most other columns, then update that row with the real data, causing the row to grow in size. Rather, insert rows filled with data from the start.
If an UPDATE statement increases the amount of data in a row so that the row no longer fits in its data block, then Oracle tries to find another block with enough free space to hold the entire row. If such a block is available, then Oracle moves the entire row to the new block. This is called migrating a row. If the row is too large to fit into any available block, then Oracle splits the row into multiple pieces and stores each piece in a separate block. This is called chaining a row. Rows can also be chained when they are inserted.
Migration and chaining are especially detrimental to performance with the following:
UPDATE statements that cause migration and chaining to perform poorlyIdentify migrated and chained rows in a table or cluster using the ANALYZE statement with the LIST CHAINED ROWS clause. This statement collects information about each migrated or chained row and places this information in a specified output table.
|
Note: Oracle Corporation strongly recommends that you use the However, you must use the |
The definition of a sample output table named CHAINED_ROWS appears in a SQL script available on your distribution medium. The common name of this script is UTLCHN1.SQL, although its exact name and location varies depending on your platform. Your output table must have the same column names, datatypes, and sizes as the CHAINED_ROWS table.
Increasing PCTFREE can help to avoid migrated rows. If you leave more free space available in the block, then the row has room to grow. You can also reorganize or re-create tables and indexes that have high deletion rates. If tables frequently have rows deleted, then data blocks can have partially free space in them. If rows are inserted and later expanded, then the inserted rows might land in blocks with deleted rows but still not have enough room to expand. Reorganizing the table ensures that the main free space is totally empty blocks.
See Also:
|
The more your application parses, the more potential for contention exists, and the more time your system spends waiting. If parse time CPU represents a large percentage of the CPU time, then time is being spent parsing instead of executing statements. If this is the case, then it is likely that the application is using literal SQL and so SQL cannot be shared, or the shared pool is poorly configured.
There are a number of statistics available to identify the extent of time spent parsing by Oracle. Query the parse related statistics from V$SYSSTAT. For example:
SELECT NAME, VALUE FROM V$SYSSTAT WHERE NAME IN ( 'parse time cpu', 'parse time elapsed' , 'parse count (hard)', 'CPU used by this session' );
There are various ratios that can be computed to assist in determining whether parsing may be a problem:
This ratio indicates how much of the CPU time spent parsing was due to the parse operation itself, rather than waiting for resources, such as latches. A ratio of one is good, indicating that the elapsed time was not spent waiting for highly contended resources.
This ratio indicates how much of the total CPU used by Oracle server processes was spent on parse-related operations. A ratio closer to zero is good, indicating that the majority of CPU is not spent on parsing.
The views V$SESSION_WAIT, V$SESSION_EVENT and V$SYSTEM_EVENT provide information on what resources were waited for, and, if the configuration parameter TIMED_STATISTICS is set to true, how long each resource was waited for.
| See Also:
"Setting the Level of Statistics Collection" for information about |
Investigate wait events and related timing data when performing reactive performance tuning. The events with the most time listed against them are often strong indications of the performance bottleneck.
The three views contain related, but different, views of the same data:
V$SESSION_WAIT is a current state view. It lists either the event currently being waited for or the event last waited for on each sessionV$SESSION_EVENT lists the cumulative history of events waited for on each session. After a session exits, the wait event statistics for that session are removed from this view.V$SYSTEM_EVENT lists the events and times waited for by the whole instance (that is, all session wait events data rolled up) since instance startup.Because V$SESSION_WAIT is a current state view, it also contains a finer-granularity of information than V$SESSION_EVENT or V$SYSTEM_EVENT. It includes additional identifying data for the current event in three parameter columns: P1, P2, and P3.
For example, V$SESSION_EVENT can show that session 124 (SID=124) had many waits on the db file scattered read event, but it does not show which file and block number. However, V$SESSION_WAIT shows the file number in P1, the block number read in P2, and the number of blocks read in P3 (P1 and P2 let you determine for which segments the wait event is occurring).
This chapter concentrates on examples using V$SESSION_WAIT. However, Oracle recommends capturing performance data over an interval and keeping this data for performance and capacity analysis. This form of rollup data is queried from the V$SYSTEM_EVENT view by tools such as Enterprise Manager Diagnostics Pack and Statspack.
Most commonly encountered events are described in this chapter, listed in case-sensitive alphabetical order. Other event-related data to examine is also included. The case used for each event name is that which appears in the V$SYSTEM_EVENT view.
| See Also:
Oracle9i Database Reference for a complete list of wait events |
The following events signify that the database process is waiting for acknowledgment from a database link or a client process:
If these waits constitute a significant portion of the wait time on the system or for a user experiencing response time issues, then the network or the middle-tier could be a bottleneck.
Events that are client-related should be diagnosed as described for the event SQL*Net message from client. Events that are dblink-related should be diagnosed as described for the event SQL*Net message from dblink.
Although this is an idle event, it is important to explain when this event can be used to diagnose what is not the problem. This event indicates that a server process is waiting for work from the client process. However, there are several situations where this event could accrue most of the wait time for a user experiencing poor response time. The cause could be either a network bottleneck or a resource bottleneck on the client process.
A network bottleneck can occur if the application causes a lot of traffic between server and client and the network latency (time for a round-trip) is high. Symptoms include the following:
To alleviate network bottlenecks, try the following:
VSAT links).If the client process is using most of the resources, then there is nothing that can be done in the database. Symptoms include the following:
In some cases, you can see the wait time for a waiting user tracking closely with the amount of CPU used by the client process. The term client here refers to any process other than the database process (middle-tier, desktop client) in the n-tier architecture.
This event signifies that the session has sent a query to the remote node and is waiting for a response from the database link. This time could go up because of the following:
For information, see "SQL*Net message from client".
It is useful to see the query being run on the remote node. Login to the remote database, find the session created by the database link, and examine the SQL statement being run by it.
This wait indicates that there are some buffers in the buffer cache that multiple processes are attempting to access concurrently. Query V$WAITSTAT for the wait statistics for each class of buffer. Common buffer classes that have buffer busy waits include data block, segment header, undo header, and undo block.
Check the following V$SESSION_WAIT parameter columns:
To determine the possible causes, identify the type of class contended for by querying V$WAITSTAT:
SELECT class, count FROM V$WAITSTAT WHERE count > 0 ORDER BY count DESC;
Example output:
CLASS COUNT ------------------ ---------- data block 43383 undo header 10680 undo block 5237 segment header 785
To identify the segment and segment type contended for, query DBA_EXTENTS using the values for File Id and Block Id returned from V$SESSION_WAIT (p1 and p2 columns):
SELECT segment_owner, segment_name FROM DBA_EXTENTS WHERE file_id = <&p1> AND <&p2> BETWEEN block_id AND block_id + blocks - 1;
The action required depends on the class of block contended for and the actual segment.
If the contention is on the segment header, then this is most likely freelist contention.
Automatic segment-space management in locally managed tablespaces eliminates the need to specify the PCTUSED, FREELISTS, and FREELIST GROUPS parameters. If possible, switch from manual space management to automatic segment-space management.
The following information is relevant if you are unable to use automatic segment-space management (for example, because the tablespace uses dictionary space management).
A freelist is a list of free data blocks that usually includes blocks existing in a number of different extents within the segment. Blocks in freelists contain free space greater than PCTFREE. This is the percentage of a block to be reserved for updates to existing rows. In general, blocks included in process freelists for a database object must satisfy the PCTFREE and PCTUSED constraints. Specify the number of process freelists with the FREELISTS parameter. The default value of FREELISTS is one. The maximum value depends on the data block size.
To find the current setting for freelists for that segment, run the following:
SELECT SEGMENT_NAME, FREELISTS FROM DBA_SEGMENTS WHERE SEGMENT_NAME = segment name AND SEGMENT_TYPE = segment type;
Set freelists, or increase of number of freelists. If adding more freelists does not alleviate the problem, then use freelist groups (even in single instance this can make a difference). If using Oracle Real Application Clusters, then ensure that each instance has its own freelist group(s).
See Also:
|
If the contention is on tables or indexes (not the segment header):
For contention on rollback segment header:
For contention on rollback segment block:
This event signifies that the user process is reading buffers into the SGA buffer cache and is waiting for a physical I/O call to return. A db file scattered read issues a scatter-read to read the data into multiple discontinuous memory locations. A scattered read is usually a multiblock read. It can occur for a fast full scan (of an index) in addition to a full table scan.
The db file scattered read wait event identifies that a full table scan is occurring. When performing a full table scan into the buffer cache, the blocks read are read into memory locations that are not physically adjacent to each other. Such reads are called scattered read calls, because the blocks are scattered throughout memory. This is why the corresponding wait event is called 'db file scattered read'. Multiblock (up to DB_FILE_MULTIBLOCK_READ_COUNT blocks) reads due to full table scans into the buffer cache show up as waits for 'db file scattered read'.
Check the following V$SESSION_WAIT parameter columns:
P1 - The absolute file numberP2 - The block being readP3 - The number of blocks (should be greater than 1)On a healthy system, physical read waits should be the biggest waits after the idle waits. However, also consider whether there are direct read waits (signifying full table scans with parallel query) or db file scattered read waits on an operational (OLTP) system that should be doing small indexed accesses.
Other things that could indicate excessive I/O load on the system include the following:
There are several ways to handle excessive I/O waits. In the order of effectiveness, these are as follows:
The first course of action should be to find opportunities to reduce I/O. Examine the SQL statements being run by sessions waiting for these events, as well as statements causing high physical I/Os from V$SQLAREA. Factors that can adversely affect the execution plans causing excessive I/O include the following:
Besides reducing I/O, also examine the I/O distribution of files across the disks. Is I/O distributed uniformly across the disks, or are there hot spots on some disks? Are the number of disks sufficient to meet the I/O needs of the database?
See the total I/O operations (reads and writes) by the database, and compare those with the number of disks used. Remember to include the I/O activity of LGWR and ARCH processes.
Use the following query to see, at a point in time, which sessions are waiting for I/O:
SELECT s.sql_address, s.sql_hash_value FROM V$SESSION s, V$SESSION_WAIT w WHERE w.event LIKE 'db file%read' AND w.sid = s.sid ;
Use the following query to find the object being accessed:
SELECT segment_owner, segment_name FROM DBA_EXTENTS WHERE file_id = &p1 AND &p2 between block_id AND block_id + blocks - 1 ;
This event signifies that the user process is reading buffers into the SGA buffer cache and is waiting for a physical I/O call to return. This call differs from a scattered read, because a sequential read is reading data into contiguous memory space. A sequential read is usually a single-block read.
Single block I/Os are usually the result of using indexes. Rarely, full table scan calls could get truncated to a single block call due to extent boundaries, or buffers already present in the buffer cache. These waits would also show up as 'db file sequential read'.
Check the following V$SESSION_WAIT parameter columns:
P1 - The absolute file numberP2 - The block being readP3 - The number of blocks (should be 1)
| See Also:
"db file scattered read" for information on managing excessive I/O, inadequate I/O distribution, and finding the SQL causing the I/O and the segment the I/O is performed on |
On a healthy system, physical read waits should be the biggest waits after the idle waits. However, also consider whether there are db file sequential reads on a large data warehouse that should be seeing mostly full table scans with parallel query.
Figure 22-1 depicts the differences between the following wait events:
db file sequential read (single block read into one SGA buffer)db file scattered read (multiblock read into many discontinuous SGA buffers)direct read (single or multiblock read into the PGA, bypassing the SGA)
When a session is reading buffers from disk directly into the PGA (opposed to the buffer cache in SGA), it waits on this event. If the I/O subsystem does not support asynchronous I/Os, then each wait corresponds to a physical read request.
If the I/O subsystem supports asynchronous I/O, then the process is able to overlap issuing read requests with processing the blocks already existing in the PGA. When the process attempts to access a block in the PGA that has not yet been read from disk, it then issues a wait call and updates the statistics for this event. Hence, the number of waits is not necessarily the same as the number of read requests (unlike 'db file scattered read' and 'db files sequential read').
Check the following V$SESSION_WAIT parameter columns:
P1 - File_id for the read callP2 - Start block_id for the read callP3 - Number of blocks in the read callThis happens in the following situations:
The file_id shows if the reads are for an object in TEMP tablespace (sorts to disk) or full table scans by parallel slaves. This is the biggest wait for large data warehouse sites. However, if the workload is not a DSS workload, then examine why this is happening.
Examine the SQL statement currently being run by the session experiencing waits to see what is causing the sorts. Query V$TEMPSEG_USAGE to find the SQL statement that is generating the sort. Also query the statistics from V$SESSTAT for the session to determine the size of the sort. See if it is possible to reduce the sorting by tuning the SQL statement. If WORKAREA_SIZE_POLICY is MANUAL, then consider increasing the SORT_AREA_SIZE for the system (if the sorts are not too big) or for individual processes. If WORKAREA_SIZE_POLICY is AUTO, then investigate whether to increase PGA_AGGREGATE_TARGET.
If tables are defined with a high degree of parallelism, then this could skew the optimizer to use full table scans with parallel slaves. Check the object being read into using the direct path reads, as well as the SQL statement being run by the query-coordinator. If the full table scans are a valid part of the workload, then ensure that the I/O subsystem is sized adequately for the degree of parallelism.
For query plans that call for a hash join, excessive I/O could result from having HASH_AREA_SIZE too small. If WORKAREA_SIZE_POLICY is MANUAL, then consider increasing the HASH_AREA_SIZE for the system or for individual processes. If WORKAREA_SIZE_POLICY is AUTO, then investigate whether to increase PGA_AGGREGATE_TARGET.
When a process is writing buffers directly from PGA (as opposed to the DBWR writing them from the buffer cache), the process waits on this event for the write call to complete. Operations that could perform direct path writes include when a sort goes to disk, during parallel DML operations, direct-path INSERTs, parallel create table as select, and some LOB operations.
Like direct path reads, the number of waits is not the same as number of write calls issued if the I/O subsystem supports asynchronous writes. The session waits if it has processed all buffers in the PGA and is unable to continue work until an I/O request completes.
Check the following V$SESSION_WAIT parameter columns:
P1 - File_id for the write callP2 - Start block_id for the write callP3 - Number of blocks in the write callThis happen in the following situations:
For large sorts see "Sorts to Disk".
For parallel DML, check the I/O distribution across disks and make sure that the I/O subsystem is adequately sized for the degree of parallelism.
Enqueues are locks that serialize access to database resources. This event indicates that the session is waiting for a lock that is held by another session.
Check the following V$SESSION_WAIT parameter columns:
P1 - Lock TYPE (or name) and MODEP2 - Resource identifier ID1 for the lockP3 - Resource identifier ID2 for the lockCheck the comparison with V$LOCK columns:
Performing the following SQL transformation of the P1 column results in the same value displayed in V$LOCK.TYPE:
V$LOCK.TYPE = chr(bitand(P1,-16777216)/16777215)|| chr(bitand(P1,16711680)/65535)
To obtain the mode in which the enqueue is being requested, issue the following statement:
request = mod(P1, 65536);
Query V$LOCK to find the sessions holding the lock. For every session waiting for the event enqueue, there is a row in V$LOCK with REQUEST <> 0. Therefore, use either of the following two queries to find the sessions holding the locks and waiting for the locks.
If there are enqueue waits, you can see these using the following statement:
SELECT * FROM V$LOCK WHERE request > 0:
To show only holders and waiters for locks being waited on, use the following:
SELECT DECODE(request,0,'Holder: ','Waiter: ')|| sid sess, id1, id2, lmode, request, type FROM V$LOCK WHERE (id1, id2, type) IN (SELECT id1, id2, type FROM V$LOCK WHERE request>0) ORDER BY id1, request;
See Also:
|
The appropriate action depends on the type of enqueue.
If the contended-for enqueue is the ST enqueue, then the problem is most likely to be dynamic space allocation. Oracle dynamically allocates an extent to a segment when there is no more free space available in the segment. This enqueue is only used for dictionary managed tablespaces.
To solve contention on this resource:
TEMPFILES. If not, then switch to using TEMPFILES.| See Also:
Oracle9i Database Concepts for detailed information on |
EXTENTS column of the DBA_SEGMENTS view for all SEGMENT_NAMEs over time to identify which segments are growing and how quickly.ALTER TABLE ALLOCATE EXTENT SQL statement).The HW enqueue is used to serialize the allocation of space beyond the high-water mark of a segment.
V$SESSION_WAIT.P2 / V$LOCK.ID1 is the tablespace number.V$SESSION_WAIT.P2 / V$LOCK.ID2 is the relative dba of segment header of the object for which space is being allocated.If this is a point of contention for an object, then manual allocation of extents solves the problem.
The most common reason for waits on TM locks tend to involve foreign key constraints where the constrained columns are not indexed. Index the foreign key columns to avoid this problem.
These are acquired exclusive when a transaction initiates its first change and held until the transaction does a COMMIT or ROLLBACK.