Release 1 (9.0.1)
Part Number A87503-02
 Home |
 Book List |
 Contents |
 Index |  Master Index |  Feedback |
| Oracle9i Database Performance Guide and Reference Release 1 (9.0.1) Part Number A87503-02 |
|
This chapter discusses the method used for performing tuning. 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.
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 below.
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.
Below are the main steps in the Oracle Performance Method:
Oracle9i Database Performance Methods for a list of common errors and for a theoretical description of this performance method
See Also:
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 per 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 it's 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 per session. Use this view to determine which particular session is using the most CPU.
V$RSRC_CONSUMER_GROUP shows CPU utilization statistics on a per consumer group basis when the Oracle Database Resource Manager is running.
It is important to recognize that CPU time and real time are distinct. With eight CPUs, for any given minute in real time, there are eight minutes of CPU time available. On NT and UNIX, this can be either user time or system time (privileged mode on NT). Thus, CPU time utilized by all processes (threads) on the system could be greater than one minute per one minute real time interval.
At any given moment, you know how much time Oracle has used on the system. So, if eight minutes are available and Oracle uses four minutes of that time, then you know that 50% of all CPU time is used by Oracle. If your process is not consuming that time, then some other process is. Identify the processes that are using CPU time, figure out why, and then attempt to tune them.
If the CPU usage is evenly distributed over many Oracle server processes, then examine the Statspack report for other evidence.
An overly active I/O system can be evidenced by disk queue lengths greater than two, or disk service times that are over 20-30ms. If the I/O system is overly active, then check for potential hot spots that could benefit from distributing the I/O across more disks. Also identify whether the load can be reduced by lowering the resource requirements of the programs using those resources.
Use operating system monitoring tools to determine what processes are running on the system as a whole and to monitor disk access to all files. Remember that disks holding datafiles and redo log files can also hold files that are not related to Oracle. Reduce any heavy access to disks that contain database files. Access to non-Oracle files can be monitored only through operating system facilities, rather than through the V$ 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$SYSTEM_EVENT to see whether the top wait events are I/O related. I/O related events include db file sequential read, db file scattered read, db file single write, and db file parallel write. These are all events corresponding to I/Os performed against the data file headers, control files, or data files. If any of these wait events correspond to high average time, then investigate the I/O contention.
Cross reference the host I/O system data with the I/O sections in the Statspack report to identify hot datafiles and tablespaces. Also compare the I/O times reported by the OS with the times reported by Oracle to see if they are consistent.
Before investigating whether the I/O system should be reconfigured, determine if the load on the I/O system can be reduced. To reduce Oracle I/O load, look at SQL statements that perform many physical reads by querying the V$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.
Using OS 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) per request.
|
See Also:
Oracle9i Database Performance Methods for a description of important operating system statistics |
Oracle statistics are examined and cross-referenced with OS statistics to ensure a consistent diagnosis of the problem. OS 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.
Below are 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$SYSTEM_EVENT can be queried for wait event statistics. If the configuration parameter TIMED_STATISTICS is set to true, then you can also see how long each resource was waited for. To minimize user response time, reduce the time spent by server processes waiting for event completion. Not all wait events have the same wait time. Therefore, it is more important to examine events with the most total time waited rather than wait events with a high number of occurrences. Usually, it is best to set the dynamic parameter TIMED_STATISTICS to true at least while monitoring performance.
Investigate wait events and related timing data when performing reactive performance tuning. The events with the most time listed against them are often strong indications of the performance bottleneck. For example, by looking at V$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.
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 on a per-file basis, including the number of I/Os per file and the average read time.
This contains detailed rollback and undo segment statistics on a per-segment basis.
This contains detailed enqueue statistics on a per-enqueue basis, 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 on a per-latch basis, 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 |
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 OS upgrade if SQL is tuned, the OS 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 per second and per transaction.
In the Statspack report, look at the Load Profile section. (The data has been normalized per transaction and per second.)
The per second statistics show the changes in throughput (that is, whether the instance is performing more work per second). The per transaction statistics identify changes in the application characteristics by comparing these to the corresponding statistics from the baseline report.
Examine the per second statistics 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.
Below 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$EVENT_NAME for a list of all wait events.) Some of these events are termed idle events, because the process is idle, waiting for work to perform. Nonidle events indicate nonproductive time spent waiting for a resource or action to complete.
|
Note: Not all symptoms can be evidenced by wait events. See "Additional Statistics" for the statistics that can be checked. |
The most effective way to use wait event data is to order the events by the wait time. This is only possible if TIMED_STATISTICS is set to true. Otherwise, the wait events can only be ranked by the number of times waited, which is often not the ordering that best represents the problem.
To get an indication of where time is spent, follow these steps:
V$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
.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 example above, 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 all 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$SYSSTAT, OS statistics, and so on). Perform cross-checks with other data to confirm or refute the developing theory.
The table below links wait events to possible causes and gives an overview of the Oracle data that could be most useful to review next.
|
See Also:
|
There are a number of statistics that can indicate performance problems that do not have corresponding wait events.
The V$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. This might not be relevant for segments that store LOBs, because these by nature can exceed the size of a block.
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 perform poorly.
Identify 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.
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.
Increase PCTFREE 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 with high deletion rates.
|
See Also:
|
The more your application parses, the more 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$SYSSTAT WHERE NAME IN ( 'parse time cpu', 'parse time elapsed' , 'parse count (hard)', 'CPU used by this session' );
There are various ratios that can be computed to assist in determining whether parsing may be a problem:
The views V$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.
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$SESSION_WAIT is a current state view. It lists either the event currently being waited for or the event last waited for on each session
V$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$SYSTEM_EVENT view.
The following events signify that the database process is waiting for acknowledgment from a database link or a client process:
If these waits constitute a significant portion of the wait time on the system or for a user experiencing response time issues, then the network or the middle-tier could be a bottleneck.
Events that are client-related should be diagnosed as described for the event SQL*Net message from client. Events that are dblink-related should be diagnosed as described for the event SQL*Net message from dblink.
Although this is an idle event, it is important to explain when this event can be used to diagnose what is not the problem. When a server process is waiting for work from the client process, it waits on this event. However, there are several situations where this event could accrue most of the wait time for a user experiencing poor response time.
This could be the case if the application causes a lot of traffic between the server and the client, and the network latency (time for a round-trip) is high. Symptoms include the following:
To alleviate network bottlenecks, try the following:
If the client process is using most of the resources, then there is nothing that can be done in the database. Symptoms include the following:
In some cases, you can see the wait time for a waiting user tracking closely with the amount of CPU used by the client process. The term client here refers to any process other than the database process (middle-tier, desktop client) in the n-tier architecture.
This event signifies that the session has sent a query to the remote node and is waiting for a response from the database link. This time could go up because of the following:
See Also:
This wait indicates that there are some buffers in the buffer cache that multiple processes are attempting to access concurrently. Query V$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.
P1 - File ID
P2 - Block ID
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.
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'.
P1 - The absolute file number
P2 - The block being read
P3 - 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 the following:
db file scattered read waits on an operational (OLTP) system that should be doing small indexed accesses?
db file sequential reads on a large data warehouse that should be seeing mostly full table scans with parallel query?
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 is 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 find the SQL statement for sessions waiting for I/O:
SELECT 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'.
P1 - The absolute file number
P2 - The block being read
P3 - 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 |
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').
P1 - File_id for the read call
P2 - Start block_id for the read call
P3 - Number of blocks in the read call
This 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$SORT_USAGE to find the SQL statement that is generating the sort. Also query the statistics from V$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.
P1 - File_id for the write call
P2 - Start block_id for the write call
P3 - Number of blocks in the write call
This 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.
P1 - Name and type of the lock
P2 - Resource identifier ID1 for the lock
P3 - Resource identifier ID2 for the lock
V$LOCK.ID1 = P2
V$LOCK.ID2 = P3
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)
The appropriate action depends on the type of enqueue.
If the contended-for enqueue is the ST enqueue, then the problem is most likely 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.
Oracle9i Database Concepts for detailed information on
See Also:
TEMPFILEs and locally managed tablespaces
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).
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 two queries to find the sessions holding the locks and waiting for the locks.
SELECT DECODE(l.request,0,'Holder: ','Waiter: ')||sid sess , id1, id2, lmode, request, type FROM V$LOCK l WHERE (l.id1,l.id2,l.type) IN ( SELECT w.p2, w.p3, chr(bitand(w.p1,-16777216)/16777215) || chr(bitand(w.p1,16711680)/65535) FROM V$SESSION_WAIT w WHERE w.wait_time = 0 AND w.event = 'enqueue' ) ORDER BY l.id1, l.request; 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 lmode = 0) ORDER BY id1, request;
|
See Also:
|
The HW enqueue is used to serialize the allocation of space above 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.
The solution is to have the first session already holding the lock perform a COMMIT or ROLLBACK.
The solution is to increase the number of ITLs available, either by changing the INITTRANS or MAXTRANS for the table (either by using an ALTER statement, or by recreating the table with the higher values).
UNIQUE index. If two sessions try to insert the same key value the second session has to wait to see if an ORA-0001 should be raised or not.
The solution is to have the first session already holding the lock perform a COMMIT or ROLLBACK.
COMMIT or ROLLBACK by waiting for the TX lock in mode 4.
This wait event indicates that a server process was unable to find a free buffer and has posted the database writer to make free buffers by writing out dirty buffers. A dirty buffer is a buffer whose contents have been modified. Dirty buffers are freed for reuse when DBWR has written the blocks to disk.
DBWR may not be keeping up with writing dirty buffers in the following situations:
If this event occurs frequently, then examine the session waits for DBWR to see whether there is anything delaying DBWR.
If it is waiting for writes, then determine what is delaying the writes and fix it. Check the following:
V$FILESTAT to see where most of the writes are happening.
If I/O is slow:
It is possible DBWR is very active because of the cache is too small. Investigate whether this is a probable cause by looking to see if the buffer cache hit ratio is low. Also use the V$DB_CACHE_ADVICE view to determine whether a larger cache size would be advantageous.
If the cache size is adequate and the I/O is already evenly spread, then you can potentially modify the behavior of DBWR by using asynchronous I/O or by using multiple database writers.
Configuring multiple database writer processes, or using I/O slaves, is useful when the transaction rates are high or when the buffer cache size is so large that a single DBWn process cannot keep up with the load.
The DB_WRITER_PROCESSES initialization parameter lets you configure multiple database writer processes (from DBW0 to DBW9). Configuring multiple DBWR processes distributes the work required to identify buffers to be written, and it also distributes the I/O load over these processes.
If it is not practical to use multiple DBWR processes, then Oracle provides a facility whereby the I/O load can be distributed over multiple slave processes. The DBWR process is the only process that scans the buffer cache LRU list for blocks to be written out. However, the I/O for those blocks is performed by the I/O slaves. The number of I/O slaves is determined by the parameter DBWR_IO_SLAVES. I/O slaves are also useful when asynchronous I/O is not available, because the multiple I/O slaves simulate nonblocking, asynchronous requests by freeing DBWR to continue identifying blocks in the cache to be written.
DBWR I/O slaves are allocated immediately following database open when the first I/O request is made. The DBWR continues to perform all of the DBWR-related work, apart from performing I/O. I/O slaves simply perform the I/O on behalf of DBWR. The writing of the batch is parallelized between the I/O slaves.
Configuring multiple DBWR processes benefits performance when a single DBWR process is unable to keep up with the required workload. However, before configuring multiple DBWR processes, check whether asynchronous I/O is available and configured on the system. If the system supports asynchronous I/O but it is not currently used, then enable asynchronous I/O to see if this alleviates the problem. If the system does not support asynchronous I/O, or if asynchronous I/O is already configured and there is still a DBWR bottleneck, then configure multiple DBWR processes.
Using multiple DBWRs parallelizes the gathering and writing of buffers. Therefore, multiple DBWn processes should deliver more throughput than one DBWR process with the same number of I/O slaves. For this reason, the use of I/O slaves has been deprecated in favor of multiple DBWR processes. I/O slaves should only be used if multiple DBWR processes cannot be configured.
A latch is a low-level internal lock used by Oracle to protect memory structures. The latch free event is updated when a server process attempts to get a latch, and the latch is unavailable on the first attempt.
This event should only be a concern if latch waits are a significant portion of the wait time on the system as a whole, or for individual users experiencing problems.
P1 - Address of the latch
P2 - Latch number
P3 - Number of times process has already slept, waiting for the latch
SELECT n.name, SUM(w.p3) Sleeps FROM V$SESSION_WAIT w, V$LATCHNAME n WHERE w.event = `latch free' AND w.p2 = n.latch# GROUP BY n.name;
A main cause of shared pool or library cache latch contention is parsing. There are a number of techniques that can be used to identify unnecessary parsing and a number of types of unnecessary parsing:
This method identifies similar SQL statements that could be shared if literals were replaced with bind variables. The idea is to either:
SELECT sql_text FROM V$SQLAREA WHERE executions = 1 ORDER BY sql_text;
SELECT SUBSTR(sql_text,1, 60), COUNT(*) FROM V$SQLAREA WHERE executions = 1 GROUP BY SUBSTR(sql_text, 1, 60) HAVING COUNT(*) > 1;
check the V$SQLAREA view. Enter the following query:
SELECT SQL_TEXT, PARSE_CALLS, EXECUTIONS FROM V$SQLAREA ORDER BY PARSE_CALLS;
When the PARSE_CALLS value is close to the EXECUTIONS value for a given statement, you might be continually reparsing that statement. Tune the statements with the higher numbers of parse calls.
Identify unnecessary parse calls by identifying the session in which they occur. It might be that particular batch programs or certain types of applications do most of the reparsing. To do this, run the following query:
column sid format 99999 column name format a20 SELECT ss.sid, sn.name, ss.value FROM V$SESSTAT ss , V$STATNAME sn WHERE name IN ('parse count (hard)','execute count') AND ss.statistic# = sn.statistic# AND ss.value > 0 ORDER BY value, sid;
The result is a list of all sessions and the amount of reparsing they do. For each system identifier (SID), go to V$SESSION to find the name of the program that causes the reparsing. The output is similar to the following:
SID NAME VALUE ------ -------------------- ---------- 7 parse count (hard) 1 8 parse count (hard) 3 7 execute count 20 6 parse count (hard) 26 11 parse count (hard) 84 6 execute count 325 11 execute count 1619 8 execute count 12690
The cache buffer lru chain latches protect the lists of buffers in the cache. When adding, moving, or removing a buffer from a list, a latch must be obtained.
For symmetric multiprocessor (SMP) systems, Oracle automatically sets the number of LRU latches to a value equal to one half the number of CPUs on the system. For non-SMP systems, one LRU latch is sufficient.
Contention for the LRU latch can impede performance on SMP machines with a large number of CPUs. LRU latch contention is detected by querying V$LATCH, V$SESSION_EVENT, and V$SYSTEM_EVENT. To avoid contention, consider bypassing the buffer cache or redesigning the application.
The cache buffers chains latches are used to protect a buffer list in the buffer cache. These latches are used when searching for, adding, or removing a buffer from the buffer cache. Contention on this latch usually means that there is a block that is greatly contended for (that is, 'hot)' block.
To identify the heavily accessed buffer chain, and hence the contended for block, look at latch statistics for the cache buffers chains latches using the view V$LATCH_CHILDREN. If there is a specific cache buffers chains child latch that has many more GETS, MISSES, and SLEEPS when compared with the other child latches, then this is the contended for child latch.
This latch has a memory address, identified by the ADDR column. Use the value in the ADDR column joined with the V$BH view to identify the blocks protected by this latch. For example, given the address (V$LATCH_CHILDREN.ADDR) of a heavily contended latch, this queries the file and block numbers:
SELECT file#, dbablk, class, state FROM X$BH WHERE HLADDR='address of latch';
There are many blocks protected by each latch. One of these buffers will likely be the hot block. Perform this query a number of times, and identify the block that consistently appears in the output, using the combination of file number (file#) and block number (dbablk). This is most likely the hot block. After the hot block has been identified, query DBA_EXTENTS using the file number and block number, to identify the segment.
This event occurs when server processes are waiting for free space in the log buffer, because you are writing redo to the log buffer faster than LGWR can write it out.
Modify the redo log buffer size. If the size of the log buffer is already reasonable, then ensure that the disks on which the online redo logs reside do not suffer from I/O contention. The log buffer space wait event could be indicative of either disk I/O contention on the disks where the redo logs reside, or of a too-small log buffer. Check the I/O profile of the disks containing the redo logs to investigate whether the I/O system is the bottleneck. If the I/O system is not a problem, then the redo log buffer could be too small. Increase the size of the redo log buffer until this event is no longer significant.
There are two wait events commonly encountered:
In both of the events, the LGWR is unable to switch into the next online redo log, and all the commit requests wait for this event.
For the log file switch (archiving needed) event, examine why the archiver is unable to archive the logs in a timely fashion. It could be due to the following:
Depending on the nature of bottleneck, you might need to redistribute I/O or add more space to the archive destination to alleviate the problem. For the log file switch (checkpoint incomplete) event:
When a user session commits (or rolls back), the session's redo information must be flushed to the redo logfile by LGWR. The server process performing the COMMIT or ROLLBACK waits under this event for the write to the redo log to complete.
If this event's waits constitute a significant wait on the system or a significant amount of time waited by a user experiencing response time issues or on a system, then examine the time per wait.
If the average time waited is low, but the number of waits are high, then the application might be committing after every INSERT, rather than batching COMMITs. Applications can reduce the wait by committing after 50 rows, rather than every row.
If the time per wait is high, then examine the session waits for the log writer and see what it is spending most of it's time doing and waiting for. If the waits are because of slow I/O, then try the following:
COMMITs by committing every N rows, rather than every row, so that fewer log file syncs are needed.
This event is used to wait for a reply from one of the background processes.
These events indicate that the server process is waiting because it has no work. This usually implies that if there is a bottleneck, then the bottleneck is not for database resources.
The majority of the idle events should be ignored when tuning, because they do not indicate the nature of the performance bottleneck. Some idle events can be useful in indicating what the bottleneck is not. An example of this type of event is the most commonly encountered idle wait-event 'SQL Net message from client'. This and other idle events (and their categories) are listed below.
|
Note:
If Statspack is installed, then it is also possible to query the |
|
 Copyright © 1996-2001, Oracle Corporation. All Rights Reserved. |
|