Haisley - Backup And Recovery Optimization

Backup and Recovery Optimization

BACKUP AND RECOVERY OPTIMIZATION Stephan Haisley, Center Of Expertise, Oracle Corporation

INTRODUCTION The objective of this paper is to explain what factors can effect the performance of backup, restoration and recovery activities. This paper also aims to provide you something to think about within the realms of backup and recovery operations. Most database administrators (DBAs) won’t spend much time on backup and recovery performance due to hopefully never needing to restore a failed database. Backup performance is only looked at when the time window allocated to backups is too small, but with the new backup duration feature in Oracle Database 10g, it is possible to tell RMAN how much time to spend on the backup, and when this time has expired RMAN will stop and then can restart from where it left off the next time it runs. Restoration and recovery performance is often only looked at when it is too late, i.e. the database has been recovering for 6 hours now, and it still hasn’t finished, but it needs to open NOW! This paper will provide you some areas to adjust and test in order to improve backup, restoration and recovery performance. This paper assumes that the reader uses Recovery Manager (RMAN) to manage the backups of their Oracle databases, and so performance of user-managed backups (non-RMAN) will not be discussed. All tests carried out in this paper use Oracle 10.1.0.3 on Suse Linux Enterprise Server 7 with a dual Intel processor machine. This was not a high end or large server so the tests are simplistic to prove the point. Throughout this paper the amount of CPU time used for an activity is shown during testing. This value is retrieved from v$sesstat for the statistic ‘CPU used by this session’ (statistic# 12) where the session is the one created when a channel is allocated by RMAN. This shows the amount of CPU time used for reading, checking/compression and writing to the backup set.

NOTE: The test results shown in this paper are not official Oracle performance figures and are simply used to demonstrate performance characteristics of backup, restoration and recovery operations. Mileage will vary greatly depending on your own hardware, operating system, network and database configurations.

RMAN OVERVIEW RMAN was introduced with Oracle 8.0 as a tool that allows the DBA to easily to manage backup and recovery for their Oracle databases. There have been improvements with each new release of Oracle and because it is built into the RDBMS kernel RMAN it can take advantage of various database features such as corruption validation checks. For a thorough description of the RMAN architecture and features, please refer to the Oracle Database Backup and Recovery Basics guide. This paper is assuming a basic understanding of the main RMAN components and terminology, and will only discuss relevant features when appropriate.

HOW RMAN TAKES A BACKUP RMAN allows the DBA to take a backup of datafiles, control files, SPFILE and archivelog files. The backups created are stored in one of two formats: Page 1 of 18

Backup and Recovery Optimization

1. Image Copies – This is an exact copy of the file being backed up, byte for byte. The Image copies that RMAN creates can be used as replacement files of the original without using RMAN to restore them. They can only be stored on disk. Because they are created from copying the original file without modification, there is very little optimization that can be done to increase performance of creating them. For this reason Image Copies will not be discussed further in this paper. 2. Backup Sets – This is logical grouping of datafiles, archivelogs, controlfiles or SPFILE (datafiles and archivelogs cannot be combined in the same backup set). Backup sets can be created on disk or to tape using the Media Management Layer (MML), which is a software layer that provides the interface from RMAN to a tape library. This is the type of backup discussed throughout this paper. When RMAN creates a backup set it creates one or more backup pieces. These pieces contain the actual data from the datafiles, archivelogs, controlfile or SPFILE. When the backup pieces are created, the data being backed up is multiplexed together so that a single backup piece will contain data from one or more backup objects (datafiles, archivelogs, controlfile, SPFILE). Because of this, only RMAN can read the data contained in the backup pieces. During backup set creation of datafiles, RMAN will scan all the way through a datafile and will not backup the data blocks that have never been used (have never contained data). This reduces the size of the backup pieces dramatically for datafiles with large amounts of free space in them. This feature is called unused block compression. Data blocks that have contained data, and then deleted to be empty will still be backed up.

INCREMENTAL BACKUPS An incremental backup is a backup method that allows you to backup only those data blocks contained in datafiles that have changed since the last incremental backup taken. Oracle provides two (0 & 1) levels (used to be five levels before Oracle Database 10g) of incremental backups as well as two different types: 1. Differential – All datablocks modified since the last level 0 or level 1 are backed up. This is the default mode of operation. 2. Cumulative – All datablocks modified since the last level 0 are backed up.

Day of the week Sun

Mon

0

1

Tues

Wed

Thr

Fri

Sat

1

1

1

1

0

Incremental backup level

Figure 1: Differential Incremental Backups

Page 2 of 18

Backup and Recovery Optimization

Day of the week Sun

Mon

Tues

Wed

Thr

Fri

Sat

0

1

1

1

1

1

0

Incremental backup level

Figure 2: Cumulative Incremental Backups

Due the reduced volume of data backed up using a differential backup, the backups will most often be smaller and also run faster than full or level 0 backups. Cumulative backups normally take longer than differential backups due to backing up more data, but during a restoration, they will run faster because fewer incremental backups will be restored. A simple test shows the difference in backup set size and time taken using differential and cumulative backups. After a level 0 backup was taken, a series of updates were run from a script that would update a similar number of data blocks each time. An incremental backup was then taken. The update script was run a second time followed by another incremental backup. This was repeated a third time to create three incremental backups. The whole test was run twice, one using differential and one using cumulative backups. The results are shown in the table below:

Backup #

Incremental Type

Incremental Level

Size of BS (8192 blocks)

Time (secs)

0

Base level

1

CPU (secs)

0

778112

626

227.20

Differential

1

42375

312

82.93

2

Differential

1

42370

312

82.65

3

Differential

1

42369

312

82.45

0

Base Level

0

778112

628

226.09

1

Cumulative

1

42371

314

80.61

2

Cumulative

1

49605

315

83.70

3

Cumulative

1

60176

321

85.33

Table 1: Speed difference between Differential and Cumulative incremental backups

As you might expect, the speed and size of the cumulative backups grew because it is backing up all data blocks that changed since the base level 0 backup. When looking at the timing of your own backups using the ELAPSED_SECONDS column in Page 3 of 18

Backup and Recovery Optimization

v$backup_set it is important to note the amount of time spent taking the level 0 in comparison to the level 1 incrementals. If the time taking a level 1 gets close to the time for a level 0, it makes practical sense to switch to using all level 0 backups due to the speed increase when recovering the database. The same backups shown in Table 1 were used to demonstrate the difference in time it takes to restore the three differential backups and the single cumulative backup. Table 2 below shows the results:

Incremental Type

Number of BS restored

Time (secs)

CPU (secs)

Base level 0

1

626.67

210.85

Differential

3

98.67

23.00

Base level 0

1

629.33

209.21

Cumulative

1

43.00

11.05

Table 2: Performance differences during restoration

The timing of the restorations (which are carried out by RMAN when a RECOVER command is issued) for the incremental backups clearly shows that cumulative incremental backups can save a lot of time during recovery. It is up to you to check the timing of your backup and recoveries to see if using cumulative backups would help reduce recovery times. The time taken to recover/restore using RMAN backups can be seen in the ELAPSED_SECONDS column of the v$session_longops view. Note that the contents of this view are cleared when the instance is shutdown and rows may get replaced with new sessions over time.

FACTORS AFFECTING BACKUP AND RESTORATION PERFORMANCE There are a number of factors that can influence backup and restoration performance: 1. Channel Configuration 2. Size of memory buffers used in creating the backup sets 3. Speed of backup devices (tape or disk) 4. Amount of data being backed up 5. Amount of block checking features enabled 6. Use of compression The process of applying recovery to the database using archivelogs and incremental backups will be discussed in a later section.

Page 4 of 18

Backup and Recovery Optimization

1. CHANNEL CONFIGURATION A channel is the communication pathway from RMAN to the physical backup device, whether it is disk or tape. It is recommended to match up the number of channels to the number of physical devices available for the backup. This applies more specifically to tape devices. When multiple channels are allocated during the backup, each channel creates its own backup sets. Oracle tries to distribute the datafiles amongst the channels with the aim of creating evenly sized backup sets. If you only have two tape devices but you allocate three channels, depending on the MML, either one channel will always be waiting for a device to finish on another channel before it can begin creating its backup set or two backup sets will be interlaced onto the tape. Interlacing backups in this way may provide quicker backup times but will slow down restore times dramatically. It is also important to consider that the available tape devices may be requested for use for another database if it needs to be restored at the same time a backup is running. Make sure there is enough backup device bandwidth for emergency recoveries to occur at any time. By using the automatic channel parallelism option RMAN will: •

Split up the number of files to be backed up amongst the channels

•

Try and even out the disks being read containing the files amongst the channels

•

Attempt to make each backupset the same size

Automatic channel parallelism is documented in chapter 2 of the Backup and Recovery Advanced Users Guide 10g. Adjusting the filesperset parameters for channels may make differences to the speed of backups but it can make major difference with restoration times when not restoring all the datafiles. This is demonstrated with a test taking several level 0 backups with a different number of files per set. A datafile is then deleted and timing information is gathered when using each of the backup sets to restore it. The results are shown below in table 3: #Files in BS

BS Size (blocks)

Restored File size (blocks)

Time (secs)

CPU (Secs)

8

702320

97727

132

39.42

4

658221

97727

110

36.92

2

132773

97727

82

29.92

1

97730

97727

74

25.62

Table 3: Effects of filesperset on restore speed It is obvious from the results that by keeping the number of filesperset to a smaller value, the speed at which the file can be restored is decreased. If the restoration involved the entire database, having more backup sets may increase the time it takes to restore. This is due to more requests being sent to the MML. This may not be a significant performance decrease but should be monitored before permanently reducing the filesperset of your backups.

2. SIZE OF MEMORY BUFFERS When RMAN takes a backup of datafiles it must read each block into an input/reading buffer. The block is checked to make sure it needs to be backed up and then various block validation checks are made in order to detect corruptions. The block is then copied into the output/write buffers. For archivelogs each logfile block also needs to be read and validated before it is written to the backup. Write buffers are used to store the multiplexed datablocks or archive log blocks, which are then written to the backup media (disk or tape). This is shown clearly in Figure 3 below:

Page 5 of 18

Backup and Recovery Optimization

Datafiles input Buffers (4 per datafile)

Output Buffers (4 per channel)

Backup Device

Figure 3: Use of memory buffers during a backup

BACKUP READING BUFFERS Oracle has documented the size and number of disk (they will always be disk due to the Oracle datafiles residing on disk) reading buffers used during a backup in the 10g Backup and Recovery Advanced User’s Guide: Number of files per Allocated channel

Buffer size

Files ≤ 4

Each buffer = 1Mb, total buffer size for channel is up to 16Mb

4 > Files ≤ 8

Each buffer = 512k, total buffer size for channel is up to 16Mb. Numbers of buffers per file will depend on number of files.

Files > 8

Each buffer = 128k, 4 buffers per file, so each file will have 512Kb buffer Table 4: Read buffer allocation algorithm

The parameter that adjusts the number and size of read buffers allocated during the backup is the channel parameter MAXOPENFILES. This specifies the maximum number of datafiles that RMAN will concurrently open for input into the backup. The default value is min(8, #files in backupset).

Page 6 of 18

Backup and Recovery Optimization

To show how the algorithm equates to the actual buffer allocation during a backup the table below shows the buffers being split up amongst datafiles and backup sets when backing up a database using different numbers for MAXOPENFILES: MAXOPENFILES

Block size (bytes)

Buffer Size (Kb)

#Buffers per file

#files open at one time

Total Buffer Size (MB)

1 2

8192

1024

16

1

16

8192

1024

8

2

16

3

8192

1024

5

3

15

4

8192

512

8

4

16

5

8192

512

6

5

15

6

8192

512

5

6

15

7

8192

512

4

7

14

8

8192

512

4

8

16

9

8192

128

4

9

4.5

10

8192

128

4

10

5

Table 5: Disk Read Buffer Allocations for Datafile Backups The number and the size of the buffers allocated for each file (datafile and archivelogs) can be seen using the following query: SELECT set_count, device_type, type, filename, buffer_size, buffer_count, open_time,close_time FROM v$backup_async/sync_io ORDER BY set_count,type, open_time, close_time;

The algorithm that Oracle uses to allocate the read buffers seems to be adequate, but it is worth monitoring v$backup_async/sync_io to show how much memory is being allocated to your backup read buffers to make sure you don’t run into memory starvation issues. When running the tests shown in table 5, there was no change to the amount of time taken to create the backup.

BACKUP WRITING BUFFERS The backup writing buffers are sized differently for tape and disk devices: Device Type

Number of Buffers Per Channel

Buffer Size

Total of Buffers Allocated

DISK

4

1Mb

4Mb

SBT

4

256Kb

1Mb

Table 6: Default write buffer allocation The SBT (tape device) write buffers are smaller than the disk channels due to the fact that the devices write slower, and hence a bigger memory area is not required. However, it is possible to modify the size of the write buffers allocated to channels, which can possible show some performance improvements. Many modern tape devices are buffered with some sort of cache so increasing the RMAN buffers should increase performance.

Page 7 of 18

Backup and Recovery Optimization

To demonstrate this I ran 3 backups of a database with different write buffer sizes. The first backup was taken using standard buffer settings (1Mb buffer). The next two backups were taken using a small and larger write buffer size, which was altered using the command: Smaller buffer ( 32k * 4 = 128Kb buffer): configure channel device type sbt parms=’BLKSIZE=32768';

Larger buffer ( 512k * 4 = 2Mb buffer): configure channel device type sbt parms='BLKSIZE=524288’;

The following table showed the results of the backups (viewed in V$BACKUP_SYNC_IO or V$BACKUP_ASYNC_IO): Total Buffer Size (bytes)

I/O Count

I/O Time (100ths sec)

131,072

60564

6174

1,048,576 (default)

7571

5959

2,097,152

3786

5053

Table 7: I/O Rates with varying write buffer sizes Because I was using the test SBT interface, which actually creates a backup to disk using the tape layer API the backup time didn’t really differ. If I were using real tape devices, increasing the buffer would have reduced the number of I/O calls to the MML significantly and subsequently the amount of I/O time would have been reduced also. It is recommended to monitor the I/O rates being achieved and then adjust the buffer size using the BLKSIZE channel parameter to a more suitable value. Of course it’s not possible to increase the I/O rate beyond what the physical device is capable of. This is explained in more detail in a later section.

SO WHERE IS THE MEMORY ALLOCATED FROM? The memory from the read and write buffers are allocated from the PGA unless I/O slaves are being used. Each channel allocated creates a separate connection into the database and will have a separate PGA allocated. If I/O slaves are used then the memory is allocated from the shared pool because the slaves will need to communicate data between themselves. If a large pool area is allocated, using the LARGE_POOL_SIZE parameter, this will be used instead of the shared pool. It is recommended to use the large pool to reduce contention in the shared pool. I/O slaves should only be used if the operating system does not support asynchronous I/O (most do in fact support it). Oracle will automatically use asynchronous or synchronous I/O if available for backups, which can be seen in V$BACKUP_SYNC_IO and V$BACKUP_ASYNC_IO. The initialization parameters that control asynchronous I/O: tape_asynch_io disk_asynch_io If these parameters are set to TRUE, yet the backup appears in the V$BACKUP_SYNC_IO, then the operation system may not be enabled to use asynchronous I/O. This is the case on my test Linux server using the SBT channel type. If not using asynchronous I/O with tape devices due to OS limitations, it is recommended to use I/O slaves which emulates asynchronous I/O. When using tape devices, the goal is to keep the tape streaming. Page 8 of 18

Backup and Recovery Optimization

What is Tape Streaming? If there is enough data being fed to the tape fast enough, the tape drive will write continuously, always having more to write at the next position on the tape after the current write is through. The drive motors can keep spinning continuously at full write speeds, so the tape just 'streams' by the heads being written. If there is not enough data to keep the tape streaming the tape drive will normally stop when no data is received and rewind back to the place of the last block written, and then wait for new data to arrive. Constantly moving the tape forward and back reduces the life of the tape, and can also slow down backup times. With synchronous I/O the channel process will wait for the write to tape to complete before continuing and filling the read buffers. When filling the read buffers the tape device will wait for the next set of write buffer data. This prevents continuous tape streaming, and is not the optimized way for taking backups. If asynchronous I/O is not available it is possible to simulate it using I/O slaves with the initialization parameters backup_tape_io_slaves. This way, the channel process allocates the writes to the slave processes and does not wait for them to complete before refilling the buffers. I ran a full database backup using SBT both with backup_tape_io_slaves set to TRUE and FALSE. Here are the results when selecting from V$BACKUP_SYNC_IO/V$BACKUP_ASYNC_IO:

1. BACKUP_TAPE_IO_SLAVES = FALSE SQL> SELECT set_count, set_stamp, device_type, type, ELAPSED_TIME, bytes FROM v$backup_async_io WHERE type=’OUTPUT’; no rows selected.

SQL> SELECT set_count, set_stamp, device_type, type, ELAPSED_TIME, bytes FROM v$backup_sync_io WHERE type=’OUTPUT’;

SET_COUNT SET_STAMP DEVICE_TYPE TYPE ELAPSED_TIME BYTES --------- ---------- -------------- --------- ------------ ---------386 549510246 SBT_TAPE OUTPUT 61500 6386450432

2. BACKUP_TAPE_IO_SLAVES = TRUE SQL> SELECT set_count, set_stamp, device_type, type, ELAPSED_TIME, bytes, filename FROM v$backup_async_io WHERE type=’OUTPUT’; SET_COUNT SET_STAMP DEVICE_TYPE TYPE ELAPSED_TIME BYTES --------- ---------- -------------- --------- ------------ ---------388 549511025 SBT_TAPE OUTPUT 61000 6386450432

SQL> SELECT set_count, set_stamp, device_type, type, ELAPSED_TIME, bytes, filename FROM v$backup_async_io WHERE output=’OUTPUT’; no rows selected

You should be using asynchronous I/O or I/O slaves to get the best backup I/O performance.

Page 9 of 18

Backup and Recovery Optimization

3. SPEED OF BACKUP DEVICES The maximum speed at which a backup can run will be dictated by: Max Mb/Sec = min(disk read Mb/s, tape write Mb/s) It is not possible to make a backup to go faster than this, period. The first part of a backup, as already explained, involves reading each datafile/archivelog file and placing the blocks into read buffers. The maximum speed at which these buffers can be filled depends on the physical location/caching and maximum read speed of the disks. The throughput being achieved through RMAN from the disks can be monitored using the effective_bytes_per_second column in the V$BACKUP_SYNC/ASYNC_IO views, where type=’INPUT’. If this is lower than the expected read rate advertised for your disks, then you need to start investigating why this is so by looking at OS data like sar and at disk monitoring data which is particular to the disk/logical volume manager implementation. The speed at which the devices used for writing the backup can also be monitored using V$BACKUP_ASYNC/SYNC_IO, using the effective_bytes_per_second column where type=’OUTPUT’. If you are seeing a slower I/O rate than expected on the tape devices then you need to look at the MML and drive options to tune the physical tape block size (normally the larger the better), the tape compression (this always slows down backups), and to make sure you are streaming data to the drive. Increasing the number of files multiplexed into each channel may increase the tape streaming abilities, but it will reduce the performance of restoring a subset of the datafiles contained in one backupset. Most of the new tape devices are equipped with some form of memory buffer so the importance of supplying adequate data for streaming is reduced. This is something you should ask of the tape device supplier. RMAN has a feature where it is possible to slow down the speed of a backup. The reason for doing this to make sure the I/O system does not get overloaded due to the backup running. By default, RMAN uses as much I/O bandwidth it can when reading the files being backed up, but it can be limited using the RATE parameter for the channels. For example, supplying the following channel configuration command will limit the read rate from disk to 1Mb/second, which if the drive offers 5Mb/sec read times, leaves plenty of I/O for other applications to function normally: RMAN> CONFIGURE CHANNEL DEVICE TYPE sbt RATE=1M;

4. AMOUNT OF DATA BEING BACKED UP (INCREMENTAL BACKUPS) Probably one of the best ways to reduce the time to carry out the backups is to simply reduce the amount of data being backed up. This is not as stupid as it sounds and with some monitoring of the backups created, they can be adjusted to reduce the backup time significantly. Here are a few guidelines: a. If there is any static data contained in the database that never changes, or changes very infrequently, such as price lookup or definition data, then place it within its own tablespace. This tablespace can be made read-only and then backed up. When it is time to make changes to the data, put the tablespace back into read-write mode, make the changes, put back into read-only mode and take another backup of it. NOTE: When backing up tablespaces very infrequently, you must make sure the backup is not purged from the tape library based on the date at which it was created. If this happens you will no longer have the data needed for recovery. If there is a limit placed on the age of backups, then make sure that a fresh backup of the read-only data is taken before the expiration. RMAN can do this automatically when using the Optimization feature. Page 10 of 18

Backup and Recovery Optimization

b. Using differential incremental level 1 backups it is possible to see the exact number of blocks that is changing for each datafile since the last level 1 backup. By using the query listed below it is possible to identify datafiles that are not changing much, and would be candidates to being backed up less frequently. SELECT set_count, set_stamp, file#, datafile_blocks, blocks “BACKUP_BLOCKS” FROM v$backup_datafile ORDER BY set_count, file#;

c. Large datafiles that contain very little data still take time to backup due to reading the whole datafile and evaluating if each datablock should be included in the backup. To demonstrate this, the following results compare taking a backup of a 500Mb file with different free space counts: %Free space

Number of Blocks Scanned

Number of Blocks Backed Up

Time Taken (secs)

100%

64000

8

22

50%

64000

31678

31

< 0.79 %

64000

63496

49

Table 8: Backup time and datafile freespace On the small test system that was used backing up to disk, the amount of time was only 49 seconds in the worst case, but you can see that even when the file was empty, it took 45% of the worst case to backup only 8 blocks. Imagine this on a busier system, with larger datafiles and many more of them. If you have large datafiles (2Gb or greater) which contain large free space amounts, it is worth while resizing them (if there are no object extents at the end of the file) and then setting them to autoextend to prevent wasted time in scanning empty space during the backup. d. Consider using the Block Change Tracking introduced in Oracle Database 10g. This allows backups using the Fast Incremental feature. Instead of scanning an entire datafile during a backup, with block change tracking enabled, a logfile (block change tracking logfile) keeps track of changed datablocks using bitmap patterns to represent ranges of datablocks. Because Oracle doesn’t keep track of all changed blocks, just a block range represented by one bit, the size of the logfile is kept very small in comparison to the size of the database. The performance overhead to updates on the database is also small. To see the difference in DML performance I used a test similar to the TPC-C benchmark using 5 warehouses and 10 concurrent users, with no thinking or keying in time. Each session carries out 5000 transactions. The test was run using a tool called HAMMERORA available at http://hammerora.sourceforge.net/tpc-c.htm. Before each test the instance was restarted and system statistics were gathered before and after the test was complete. There was no other activity on the database. Session running times were gathered using logon and logoff triggers. With block change tracking disabled the average time for a session to complete was 28mins 52secs. When I turned on block change tracking the average time increased to 29mins 49secs. This equates to a performance degradation of ~3% which is not very high considering the improvements in incremental backup speed as demonstrated in the next test. The improvement in backup speed using the fast incrementals with the block change tracking enabled was dramatic. To test this, the database was backed up using a level 0 incremental, which scans all the datafiles from start to end even if block change Page 11 of 18

Backup and Recovery Optimization

tracking is enabled. DML activity was generated in the database and then a level 1 differential incremental was taken. The exercise was repeated with block change tracking disabled and then enabled. The table below shows the results: Fast Incrementals?

#Blocks in Database

#Blocks Read

#Blocks in Backup

Time Taken (secs)

No

404160

404160

36567

156

Yes

404160

72832

37215

35

Table 9: Speed increase using Fast Incremental backups It is clear to see the increase in speed due to the reduced number of blocks being read when using the fast incremental backups. The number of blocks read should be monitored because if this rises to a value closer to the number of blocks contained in the datafile, then block change tracking will operate slower than a normal incremental level 0 due to reading the bitmap information along with the full datafile. To check if a backup used the fast incremental feature, look at the USED_CHANGE_TRACKING column in v$backup_datafile. The speed of restoration is not affected by using fast incremental backups, because the resulting backup set remains the same as if it were backed up using normal incremental backups, just created quicker.

5. AMOUNT OF BLOCK CHECKING FEATURES ENABLED RMAN offers the ability to run several different block sanity checks to detect any corruptions at time of backup and restoration. A.

HEAD AND TAIL

When RMAN is reading datafiles being backed up it is always checking that the tail portion (last 4 bytes in each block) matches the structures in the header (detecting a fractured/split block). If a mismatch is found, the block is re-read up to five times. If the same mismatch occurs on the last read, the block is reported as corrupt in the RMAN session and an error is reported in the alert.log. This type of block checking cannot be turned off. B.

CHECKSUMS

By default RMAN calculates checksums for every block read and written as part of the backup, irrespective of the db_block_checksum init.ora parameter setting. If the block already contains a checksum value, RMAN will read it and validate it against the newly calculated value. If there is a checksum mismatch the problem will be reported. The checksums calculated by RMAN are stored with the block in the backupset. On the restore the checksum is recalculated and compared, and depending on the value for db_block_checksum the checksum may get cleared before the block is written to disk. The checksum creation can be turned off by specifying the NOCHECKSUM clause as part of the backup command. However, this is not recommended. The checksums are always created for the SYSTEM datafiles and the controlfile. C.

LOGICAL BLOCK CHECKS

RMAN also offers a feature to check the logical structures contained within each data block being backed up. For example, free space counts are correct, the row index directory is valid, index block pointers are correct and no rows overlap each other. The checks are the same logical block checks as those used when the db_block_checking initialization parameter is used. To turn on logical checking during the backup use the CHECK LOGICAL keywords within the backup command. By default, logical checking is not used.

Page 12 of 18

Backup and Recovery Optimization

To demonstrate what effects on backup performance each check has, I ran a full database backup with each option enabled. The results are shown below: Checksums?

Check Logical?

Blocks read

Time (secs)

CPU (secs)

No

No

825920

623.67

227.08

Yes

No

825920

622.33

229.93

Yes

Yes

825920

624.67

245.81

Table 10: Effects of checksum and logical block checks on backups Table 10 shows that there really isn’t any difference between the calculation of checksums and without them. Each datablock already contains a checksum due to the default of init.ora value of DB_BLOCK_CHECKSUMS being set to TRUE so when RMAN is reading each block for the backup, it is computing a new checksum and comparing it against the one already contained in the datablock. It is not recommended to turn off the checksums due to the value added by this RMAN feature in block corruption detection during backups. Adding the CHECK LOGICAL parameter to the RMAN backup shows a slight decrease in performance in time and the amount of CPU used. The backups used in the test were small in size, and to disk, but it does give a guideline to the additional time and effort RMAN must make in order to check the logical structures with each block being included in the backup set. This should be tested and benchmarked in your own environment to see what the performance effects are. The effects of checksums and logical block checks are similar during RMAN restores: Checksums?

Check Logical?

#Blocks in DB

Time (secs)

CPU (secs)

No

No

825920

559

210.53

Yes

No

825920

559

211.14

Yes

Yes

825920

610

218.53

Table 11: Effects of checksum and logical block checks on restores Due to my small test system the effects of turning on checksums and logical block checks are not that great, but when backing up and restoring a multi-terabyte database, the overheads will increase. But disabling the block checking features comes at a much larger price because RMAN can no longer recognize and alert you of block corruptions. Having to restore and recover parts of the database due to corruption will take far longer than the overheads involved in backing up and restoring with the checksum and check logical options.

6. USE OF COMPRESSION Before Oracle Database 10g the only type of compression that RMAN would use was that of unused block compression – unused blocks were not backed up. In Oracle Database 10g a binary compression feature was introduced that will compress the size of the backup sets. Due the increased amount of work involved in the compression, backup and restore times will increase. CPU usage will also increase.

Page 13 of 18

Backup and Recovery Optimization

To demonstrate this, a full database backup was taken and the time, amount of CPU and final backup set size was recorded: #Blocks Read

Compression?

Time (secs)

CPU (secs)

Backup Set Size (Mb)

825920

No

622

227

6079

825920

Yes

1194

1100

517

Table 12: Effect of RMAN compression on backup speed The use of RMAN compression provides an impressive compression rate of approximately 92%. For users that have limited storage space for backups this is great, but it does come with the added CPU and time costs due to the amount of work it takes to apply a compression algorithm. If the MML also offers tape compression testing should be carried out to see if it offers a better compression ratio and backup time than using RMAN compression. You should never use compression with RMAN and the MML due to decreased performance during backup and restoration. It would be wise to test the time and CPU used for both RMAN and MML to see which offers the best value. If you don’t have any concerns about the amount of space a backup set occupies then don’t use compression. Restoring a compressed backup set has similar performance effects on the CPU usage and overall time: Compression

#Blocks Read in Backup Set

Backup Set Size (Mb)

Time (secs)

CPU (secs)

No

778109

6079

559

209.37

Yes

778109

517

1133

1093.37

Table 13: Effect of RMAN compression on restoration speed Similar to the backup compression tests, the time it takes to restore the backup is double, with a higher amount of time spent on the CPU. The use of RMAN compression would be advantageous if the tape devices are mounted across the network. If this is the case, RMAN must transfer the backup pieces from the database server to the tape server. The use of compression will significantly reduce the amount of data traveling across the network, possibly reducing the backup and restore times as well as causing less inconvenience to other network users.

Page 14 of 18

Backup and Recovery Optimization

FACTORS EFFECTING MEDIA RECOVERY PERFORMANCE? There are a number of factors that will affect the performance of recovery, including: -

Number of archivelogs and/or incrementals being applied

-

Number of datafiles needing recovery

-

If the archivelogs are available on disk outside of a backup set

-

If using parallel recovery

-

General database performance

This section is making the assumption that at this stage of recovery the datafiles have been restored from RMAN using a base level backup (level 0 or a full backup).

1. THE NUMBER OF ARCHIVELOGS AND/OR INCREMENTALS BEING APPLIED After the base level backup has been restored by RMAN, the recovery phase begins. RMAN will first look for any suitable incremental backups that can be used to roll the database forward. If no suitable incrementals backups are found, the required archive logs are looked for (assuming you are running in ARCHIVELOG mode). It has long been documented in several places that recovering the database using incremental backups is faster than using archive redo log files. This can easily be tested, by taking a base level 0 backup of the database in ARCHIVELOG mode. For a period of time run DML against the database so that it creates between 10 and 20 archivelogs, and then take an incremental backup. Delete the database and restore the base level 0 backup. Then use RMAN to recover the database, which will restore the incremental backup, timing it. Restore the base level 0 again, but instead of using RMAN to recover the database, use SQL*Plus with a ‘RECOVER AUTOMATIC DATABASE’ command, and compare the time it takes with the incremental restoration. I ran this test, and found the incremental restore to take 46 seconds compared to the recovery through applying archive logs taking 789 seconds. On the test system the backups were created on disk, so the speed of restoring incremental backups were predictably much faster. If the incremental backups were coming from a tape device or tape silo that had to locate the tape, mount the tape and then find the start of the backup, it may have taken much longer. The point is, on each system the speed at which archivelogs and restorations from the backup media are going to vary significantly, so it is worth some investment in some time to test the differences. Only then can you implement the backup strategy that will meet the time expectations when a failure occurs. As documented in an earlier test, the type of incremental backups (cumulative or differential) will also help determine if applying archivelogs are slower than restoring an incremental backup.

2. THE NUMBER OF DATAFILES NEEDING RECOVERY When media recovery is carried out, each datablock affected by a redo record must be read into the buffer cache before the redo can be applied to it. If you recover more datafiles than are necessary, you are causing recovery to do more work than necessary. Only restore and recover the minimum amount of datafiles needed to fix the failure – there is no point restoring all 100 datafiles if only 10 have got failures, unless of course you are running in NOARCHIVELOG mode when media recovery is not an option for you. If the failure is due to block corruptions within a datafile, consider using the Block Media Recovery feature introduced in Oracle9i of RMAN. Instead of restoring and recovering a whole datafile, individual datablocks can be restored from a backup, and the archivelogs and/or incremental backups are used to recover them. To show the amount of time this can save, a test that corrupted a different number of datablocks in a 102400 8Kb block datafile was carried out to show the total recovery time (including the restoration) between block media recovery and datafile recovery. The block corruptions were spread evenly throughout the datafile. Before corrupting the blocks a level 0 backup was taken of the datafile, the TPCC tests were run against the database (changing 34113 datablocks in the datafile), and all the archivelogs remained on disk so they didn’t need to be restored during recovery.

Page 15 of 18

Backup and Recovery Optimization

Number of Corrupt Blocks

Datafile Recovery Time (secs)

Block Media Recovery Time (secs)

10

941

145

99

925

155

991

937

219

5000

922

616

10000

938

1156

Table 14: Speed of block media recovery compared to datafile recovery As can be seen in Table 14, block media recovery showed a significant time saving for recovering individual blocks over restoring and recovering the whole datafile. There will be some point at which block media recovery becomes more expensive than recovering the whole datafile, and in the test this was seen somewhere around 10000 blocks which equates to approximately ~10% of the datafile. This value will be different on everybody’s system and also depending how the corrupt blocks are spread out, but this demonstrates that using BMR on too many blocks within a datafile can in fact be slower.

3. IF THE ARCHIVELOGS ARE AVAILABLE ON DISK OUTSIDE OF A BACKUP SET If the archivelogs needed for recovery are already on disk, recovery will complete quicker than if the archivelogs need to be restored from an. If backup compression (RMAN or MML) is being used recovery times will take even longer. On a simple test to apply 20 archive logs or restore them and apply them on a test system, it took over 2 minutes extra to restore them, and this was using a disk backup. This will be greatly increased if the backups are on tape and you need to apply a hundred or more archivelogs. It is a good idea to keep a number of archivelogs available on disk to aid in recovery speed should it be needed. The number of logfiles to keep depends on the frequency of backups and available disk space to store them.

4. IF USING PARALLEL RECOVERY By default Oracle uses a single process, which is the one issuing the recovery command, to carry out media recovery unless using PARALLEL_AUTOMATIC_TUNING init.ora parameter. This single process will read the archive logs and determine which changes are required for the specific recovering datafiles. The datablock is read into the buffer cache, and the redo is applied to it. Each time an archivelog is completely applied a media recovery checkpoint occurs which signals DBWR to write all the dirty blocks in the buffer cache to the datafiles. The file headers and controlfile also get updated with checkpoint information. This is a lot to do for a single process, especially when recovering a large number of datafiles applying a large amount of redo. To help decrease the time it takes to carry out media recovery Oracle provides the Parallel Recovery feature. On multiple CPU machines, you can specify a degree of parallelism for use with the RECOVER command. The process that issues the recovery command reads the archive logs as before, but instead of reading the blocks in the cache and applying the redo to them directly, it passes that work to the parallel execution slave processes. To make sure the redo is applied to the datafiles in SCN order, the parallel slaves work on different ranges of datablocks, so they will not interfere with each other, and also the redo will still get applied in SCN order for each datablock. When recovering a small number of datafiles using parallel recovery it may take longer to perform due to the extra computation involved in partitioning the work, plus the extra IPC communication between the slave and coordinator processes. Monitor v$session_wait and v$system_events for the top wait events. If you are seeing ‘PX Deq’ events, for which there are several different types, then reducing the degree of parallelism may increase the recovery time.

Page 16 of 18

Backup and Recovery Optimization

You will also notice an increase in CPU usage when using parallel recovery due to the fact that the coordinator process is calculating the work partitioning for each redo change in the archive log, and there are more processes carrying out the recovery. So long as the CPU is not being saturated (look at something equivalent to ‘sar –u’) then parallel recovery is not causing CPU issues. The final thing to note about parallel recovery is the way it uses memory. If the init.ora parameter PARALLEL_AUTOMATIC_TUNING is set to FALSE (its default value) the buffers used for messaging between the parallel processes is 2148bytes and is allocated from the shared pool. If the PARALLEL_AUTOMATIC_TUNING is set to TRUE, the default buffer is 4096bytes and allocated from the large pool. By adjusting the size of the buffers, by setting PARALLEL_EXECUTION_MESSAGE_SIZE speed of parallel recovery may be decreased, but the use of memory in the shared or large pool should also be monitored to prevent resource starvation. For more information on media recovery, take a look at the following white paper: http://www.oracle.com/technology/deploy/availability/pdf/MAA_WP_10gRecoveryBestPractices.pdf

5. GENERAL DATABASE PERFORMANCE If the database performs slowly in day-to-day activities you shouldn’t expect a fast recovery time. Recovery uses the database to carry out its tasks so if there are performance issues before recovery, recovery may have similar issues. Common areas that should be optimized that can help recovery times include: •

I/O – Recovery is very read and write intensive due to having to read all the archivelog contents, read the datablocks from the datafiles, and then write the dirty blocks backup to disk once the redo has been applied. By monitoring v$filestat along with OS specific tools, you need to make sure the read and write times are acceptable for the hardware being used. Slow I/O can significantly slow down the time it takes to carry out media recovery.

•

DBWR performance – DBWRs main responsibility is to ensure there are enough clean buffers in the buffer cache to be used when data is being read from the datafiles. When a block has been updated, it is DBWRs job to write the buffer to disk and return it for reuse within the buffer cache. If DBWR cannot keep up with cleaning the buffers, you will notice waits for the ‘free buffer waits’ event. If the I/O is not a problem, then using multiple DBWR process (DBWR slaves if your OS does not support asynchronous I/O or asynchronous I/O is disabled) should reduce the waits.

•

CPU Utilization – Because each datablock that requires recovery is read into the buffer cache before the redo change is applied to it there is a number of latches that must be acquired first. This includes the cache buffers chains and the cache buffers lru chain. Acquiring latches and making the changes to the blocks all takes CPU cycles, so you should make sure there is enough CPU bandwidth for the database during media recovery. It is not uncommon to see a server with a slow trickle of transactions with lots of CPU to spare turn into a CPU hog during media recovery due to having to apply a much more concentrated amount of datablock changes. If you are using parallel recovery CPU usage will be even higher as already discussed.

Page 17 of 18

Backup and Recovery Optimization

CONCLUSION This paper has demonstrated a number of areas to consider to optimize the speed of backups, restorations and recoveries. The speed of backup and restore activities can be controlled by: •

Using incremental backups, and cumulative incremental backups if restoration speed is more important than backup speed

•

Allocating one channel to each available tape device or controller if tape devices are shared on a single controller

•

Increasing the size of the memory buffers used when creating backups to tape

•

Reducing the amount of data that needs to be backed up – scanning empty or static datafiles can waste significant time and resources

•

Using the Block Change Tracking feature to substantially increase performance of incremental backups

•

The type of block checking features enabled for backup and restoration

•

The type of compression being used by RMAN or the Media Management Layer

The speed of media recovery can be controlled by: •

The number of archivelogs that need applying with the greater the number, the slower recovery will take

•

The number of incremental backups that need to be restored and applied

•

The type of incremental backup being used (differential or cumulative)

•

The number of datafiles needing recovery. A higher number datafiles that are restored and recovered, means a higher number of datablocks need to be read into the buffer cache and written back to the datafiles

•

Using block media recovery instead of restoring and recovering whole datafiles due to corruption issues

•

Using parallel recovery, which partitions the work of applying redo between parallel execution slave process to decrease the time it takes to apply the redo

•

General database performance tuning. If the database performs slowly, so will the recoveries, so the database should be optimized before beginning recovery optimization.

The tests carried out in this paper demonstrate the key points to optimization, and it should be understood that each system will offer different performance gains by making any of the suggested adjustments. The speed of backup, restoration and recovery is heavily dependent on the speed of the hardware being used by the system, plus any network latency introduced when the backups are stored on remote tape servers. Backup, restoration and recovery should be thoroughly tested, not just to ensure it will protect you against failures but also to make sure the backups occur in the allotted time window and the restore and recovery can complete within an agreed Service Level Agreement. Don’t leave tuning a backup, restore or recovery until it is performing slowly in production when you are in the middle of a time constraint.

Page 18 of 18