Checkpoint Tuning and Troubleshooting Guide(Metalink:147468.1)

This bulletin provides the Database Administrator a better understanding of checkpoint processing and a description of four key initialization parameters
used for checkpoint tuning: 

           -  CHECKPOINT_PROCESS
           -  LOG_CHECKPOINT_INTERVAL
           -  LOG_CHECKPOINT_TIMEOUT
           -  LOG_CHECKPOINTS_TO_ALERT 

[@more@]

It also explains how to interpret and handle checkpoint errors: 'Checkpoint not Complete' and 'Cannot Allocate New Log' reported in the ALERT.LOG file. 

Contents:  

1.  What is a Checkpoint?
2. The checkpoint process
3. Checkpoints and Performance
4. Redo logs and Checkpoint
5. Instance parameters related with the checkpoint process
6. Understanding Checkpoint Error messages    ("Cannot allocate new log" and "Checkpoint not complete")
7. Oracle Release Information 

CHECKPOINT TUNING AND ERROR HANDLING

1.  What is a Checkpoint?

A Checkpoint is a database event which synchronizes the modified data blocks in memory with the datafiles on disk.  It offers Oracle the means for ensuring the consistency of data modified by transactions.  The mechanism of writing modified blocks on disk in Oracle is not synchronized with the commit of the corresponding transactions.

A checkpoint has two purposes: (1) to establish data consistency, and (2) enable faster database recovery.   How is recovery faster?  Because all database changes up to the checkpoint have been recorded in the datafiles, making it unnecessary to apply redo log entries prior to the checkpoint. The checkpoint must ensure that all the modified buffers in the cache are really written to the corresponding datafiles to avoid the loss of data which may occur with a crash (instance or disk failure). 

Oracle writes the dirty buffers to disk only on certain conditions: 

- A shadow process must scan more than one-quarter of the db_block_buffer parameter. 
- Every three seconds. 
- When a checkpoint is produced.

A checkpoint is realized on five types of events: 

- At each switch of the redo log files. 
- When the delay for LOG_CHECKPOINT_TIMEOUT is reached. 
- When the size in bytes corresponding to : 
  (LOG_CHECKPOINT_INTERVAL* size of IO OS blocks)  is written on the current redo log file. 
-  Directly by the ALTER SYSTEM SWITCH LOGFILE command. 
- Directly with the ALTER SYSTEM CHECKPOINT command. 

During a checkpoint the following occurs:
-  The database writer (DBWR) writes all modified database blocks in the buffer cache back to datafiles,
-  Log writer (LGWR) updates both the controlfile and the datafiles to indicate when the last checkpoint occurred (SCN)

2. The checkpoint process
 
Please note that this parameter is obsolete starting with Oracle8 and now the CKPT process is always started as part of the background processes.

The CHECKPOINT_PROCESS init.ora parameter determines whether or not the optional CKPT background process will be started to perform LGWRs tasks during checkpoint operations of updating the datafile headers.  LGWR is then free to perform its' primary function flushing the redo log buffer to the online redo logs.

The CKPT process can improve performance significantly and decrease the amount of time users have to wait for a checkpoint operation to complete.  The
overhead associated with starting another background process is not significant when compared to the performance benefit to be gained by enabling
CKPT, therefore, Oracle recommends always enabling the checkpoint process (CKPT).

3. Checkpoints and Performance

Checkpoints present a tuning dilemma for the Database Administrator.  Frequent checkpoints will enable faster recovery, but can cause performance
degradation. How then should the DBA address this?

Depending on the number of datafiles in a database, a checkpoint can be a highly resource intensive operation, since all datafile headers are frozen
during the checkpoint.  There is a performance trade-off regarding frequency of checkpoints.  More frequent checkpoints enable faster database recovery
after a crash.  This is why some customer sites which have a very low tolerance for unscheduled system downtime will often choose this option. However, the performance degradation of frequent checkpoints may not justify this philosophy in many cases. Let's assume the database is up and running 95%
of the time, and unavailable 5% of the time from infrequent instance crashes or hardware failures requiring database recovery.  For most customer sites, it
makes more sense to tune for the 95% case rather than the rare 5% downtime.

This bulletin assumes that performance is your number one priority and so recommendations are made accordingly. Therefore, your goal is to minimize the frequency of checkpoints through tuning. Tuning checkpoints involves four key initialization parameters 

           -  CHECKPOINT_PROCESS
           -  LOG_CHECKPOINT_INTERVAL
           -  LOG_CHECKPOINT_TIMEOUT
           -  LOG_CHECKPOINTS_TO_ALERT

These parameters are discussed in detail below. Recommendations are also given for handling "checkpoint not complete" messages
found in the alert log, which indicate a need to tune redo logs and checkpoints.

4. Redo logs and Checkpoint

A checkpoint occurs at every log switch.  If a previous checkpoint is already in progress, the checkpoint forced by the log switch will override the current
checkpoint.

This necessitates well-sized redo logs to avoid unnecessary checkpoints as a result of frequent log switches.  The alert log is a valuable tool for
monitoring the rate that log switches occur, and subsequently, checkpoints occur.  Oracle recommends sizing the online redo logs such that switches occur
no more than once per hour. The following is an example of quick log switches from the alert log: 

Fri May 16 17:15:43 1997
Thread 1 advanced to log sequence 1272
  Current log# 3 seq# 1272 mem# 0: /prod1/oradata/logs/redologs03.log
Thread 1 advanced to log sequence 1273
  Current log# 1 seq# 1273 mem# 0: /prod1/oradata/logs/redologs01.log
Fri May 16 17:17:25 1997
Thread 1 advanced to log sequence 1274
  Current log# 2 seq# 1274 mem# 0: /prod1/oradata/logs/redologs02.log
Thread 1 advanced to log sequence 1275
  Current log# 3 seq# 1275 mem# 0: /prod1/oradata/logs/redologs03.log
Fri May 16 17:20:51 1997
Thread 1 advanced to log sequence 1276
  Current log# 1 seq# 1276 mem# 0: /prod1/oradata/logs/redologs01.log

If redo logs switch every 3 minutes, you will see performance degradation. This indicates the redo logs are not sized large enough to efficiently handle the transaction load. size of the redolog files.

5. Instance parameters related with the checkpoint process

LOG_CHECKPOINT_INTERVAL

The LOG_CHECKPOINT_INTERVAL init.ora parameter controls how often a checkpoint operation will be performed based upon the number of operating system blocks that have been written to the redo log.  If this value is larger than the size of the redo log, then the checkpoint will only occur when Oracle performs a log switch from one group to another, which is preferred.
NOTE: Starting with Oracle 8.1, LOG_CHECKPOINT_INTERVAL will be interpreted to mean that the incremental checkpoint should not lag the tail of the log by more than log_checkpoint_interval number of redo blocks.

On most Unix systems the operating system block size is 512 bytes.  This means that setting LOG_CHECKPOINT_INTERVAL to a value of 10,000 (the default
setting), causes a checkpoint to occur after 5,120,000 (5M) bytes are written to the redo log.  If the size of your redo log is 20M, you are taking 4 checkpoints for each log.

LOG_CHECKPOINT_INTERVAL influences when a checkpoint occurs, which means careful attention should be given to the setting of this parameter, keeping it updated as the size of the redo log files is changed.  The checkpoint frequency is one of the factors which impacts the time required for the database to recover from an unexpected failure.  Longer intervals between checkpoints mean that if the system crashes, more time will be needed for the database to recover.  Shorter checkpoint intervals mean that the database will recover more quickly, at the expense of increased resource utilization during the checkpoint operation.

This parameter also impacts the time required to complete a database recovery operation during the roll forward phase of recovery.  The actual recovery time is dependent upon this time, and other factors, such as the type of failure (instance or system crash, media failure, etc.), and the number of archived
redo logs which need to be applied.

LOG_CHECKPOINT_TIMEOUT

The LOG_CHECKPOINT_TIMEOUT init.ora parameter controls how often a checkpoint will be performed based on the number of seconds that have passed since the last checkpoint. 
NOTE: Starting with Oracle 8.1, LOG_CHECKPOINT_TIMEOUT will be interpreted to mean that the incremental checkpoint should be at the log position where the tail of the log was LOG_CHECKPOINT_TIMEOUT seconds ago.
Checkpoint frequency impacts the time required for the database to recover from an unexpected failure.  Longer intervals between checkpoints mean that more time will be required during database recovery. Oracle recommends using LOG_CHECKPOINT_INTERVAL to control the checkpoint interval rather than LOG_CHECKPOINT_TIMEOUT, which will initiate a checkpoint every "n" seconds, regardless of the transaction frequency.  This can cause unnecessary checkpoints in cases where transaction volumes vary.  Unnecessary checkpoints must be avoided whenever possible for optimal performance.

It is a misconception that setting LOG_CHECKPOINT_TIMEOUT to a given value will initiate a log switch at that interval, enabling a recovery window used
for a stand-by database configuration.  Log switches cause a checkpoint, but a checkpoint does not cause a log switch.  The only way to cause a log switch is
manually with ALTER SYSTEM SWITCH LOGFILE or resizing the redo logs to cause more frequent switches.  This is controlled by operating system blocks, not a timed interval.

Sizing of the online redo logs is critical for performance and recovery.

See additional sections below on redo logs and checkpoints.

LOG_CHECKPOINTS_TO_ALERT:

The LOG_CHECKPOINTS_TO_ALERT init.ora parameter, when set to a value of TRUE, allows you to log checkpoint start and stop times in the alert log.  This is very helpful in determining if checkpoints are occurring at the optimal frequency and gives a chronological view of checkpoints and other database
activities occurring in the background.

See <Note:76713.1> to have more detail on How those instance parameters can influence the checkpoint.
6. Understanding Checkpoint Error messages  (Cannot allocate new log and Checkpoint not complete)

Sometimes, you can see in your alert.log file, the following corresponding messages: 

  Thread 1 advanced to log sequence 248 
  Current log# 2 seq# 248 mem# 0: /prod1/oradata/logs/redologs02.log
  Thread 1 cannot allocate new log, sequence 249
  Checkpoint not complete 

This message indicates that Oracle wants to reuse a redo log file, but the corresponding checkpoint associated is not terminated.  In this case, Oracle
must wait until the checkpoint is completely realized. This situation may be encountered particularly when the transactional activity is important. 

 This situation may also be checked by tracing two statistics in the BSTAT/ESTAT report.txt file.  The two statistics are: 

  - Background checkpoint started.
  - Background checkpoint completed.

These two statistics must not be different more than once.  If this is not true, your database hangs on checkpoints.  LGWR is unable to continue
writing the next transactions until the checkpoints complete. Three reasons may explain this difference:

- A frequency of checkpoints which is too high.
- A checkpoints are starting but not completing
- A DBWR which writes too slowly.

The number of checkpoints completed and started as indicated by these statistics should be weighed against the duration of the bstat/estat report.  Keep in mind the goal of only one log switch per hour, which ideally should equate to one checkpoint per hour as well. The way to resolve incomplete checkpoints is through tuning checkpoints and logs:

1) Give the checkpoint process more time to cycle through the logs
       -  add more redo log groups
       -  increase the size of the redo logs
2) Reduce the frequency of checkpoints
        - increase LOG_CHECKPOINT_INTERVAL
        - increase size of online redo logs
3) Improve the efficiency of checkpoints enabling the CKPT process with CHECKPOINT_PROCESS=TRUE
4) Set LOG_CHECKPOINT_TIMEOUT = 0.  This disables the checkpointing based on time interval.
5) Another means of solving this error is for DBWR to quickly write the dirty buffers on disk.  The parameter linked to this task is: 

 DB_BLOCK_CHECKPOINT_BATCH. 
DB_BLOCK_CHECKPOINT_BATCH specifies the number of blocks which are dedicated inside the batch size for writing checkpoints.  When you want to accelerate the checkpoints, it is necessary to increase this value.

7. Oracle Release Information

The CKPT process is optional in lower versions of Oracle7, but is mandatory in Oracle8. In versions 7.0 - 7.3.2, the CKPT is an optional background process which is enabled by setting CHECKPOINT_PROCESS=TRUE in init.ora. In versions 7.3.3 and 7.3.4, the CKPT process will be started automatically regardless of the CHECKPOINT_PROCESS setting if either of the following conditions exist: 

 - a large value for DB_FILES (50 or higher)
 - a large value for DB_BLOCK_BUFFERS (10,000 or higher)

In version 8.0.3 and higher, the CKPT process is always enabled.  Attempting to set CHECKPOINT_PROCESS in the init.ora will give the following error: 

 LM-101 "unknown parameter name checkpoint_process"

Starting from Oracle8i, Oracle Corporation recommends that Enterprise Edition users who were using incremental checkpoints in an earlier release to use fast-start checkpointing in Oracle8i. In fast-start checkpointing, the FAST_START_IO_TARGET parameter replaces DB_FILE_MAX_DIRTY_TARGET.
FAST_START_IO_TARGET specifies the number of I/Os that should be needed during crash or instance recovery.When you set this parameter, DBWR writes dirty buffers out more aggressively, so that the number of blocks that must be processed during recovery stays below the value specified in the parameter.

So in Oracle8i The incremental checkpoint position should not lag the tail of the log by more than LOG_CHECKPOINT_INTERVAL operating system blocks.The
LOG_CHECKPOINT_INTERVAL and FAST_START_IO_TARGET is taken into account to determine how far behind the end of the redo stream the checkpoint position can really be.

In Oracle9i FAST_START_MTTR_TARGET is the preferred method of specifying how far the checkpoint position should be behind the tail of the redo stream. However, LOG_CHECKPOINT_INTERVAL is still supported if needed. It functions as per the Oracle8i behaviour above.

FAST_START_MTTR_TARGET enables you to specify the number of seconds the database takes to perform crash recovery of a single instance. AST_START_MTTR_TARGET can be overridden by either FAST_START_IO_TARGET or LOG_CHECKPOINT_INTERVAL.

8. Using Statspack to determine Checkpointing problems
Statspack snapshots can be taken every 15 minutes or so, these reports gather useful information about number of checkpoints started and checkpoints completed and number of database buffers written during checkpointing for that window of time . It also contains statistics about redo activity. Gathering and comparing these snapshot reports gives you a complete idea about checkpointing performance at different periods of time.

Another important thing to watch in statspack report is the following wait events, they could be a good indication about problems with the redo log throughput and checkpointing:
log file switch (checkpoint incomplete)
log file switch (archiving needed)
log file switch/archive
log file switch (clearing log file)
log file switch completion
log switch/archive
log file sync
In the case when one or more of the above wait events is repeated frequently with considerable values then you need to take an action like adding More
online redo log files or increasing their sizes and/or modifying checkpointing parameters.

来自 “ ITPUB博客 ” ，链接：http://blog.itpub.net/60857/viewspace-778849/，如需转载，请注明出处，否则将追究法律责任。