（轉）關於 db block gets,consistent gets,physical reads的概念

看到這篇文章，對於這3個概念感覺還是理解不足，轉載過來以備研究檢視

在Oracle的文件中有這樣的解釋：
Recursive Calls: Number of recursive calls generated at both the user and system level.
Oracle Database maintains tables used for internal processing. When it needs to change these tables, Oracle Database generates an internal SQL statement, which in turn generates a recursive call.
In short, recursive calls are basically SQL performed on behalf of your SQL. So, if you had to parse the query, for example, you might have had to run some other queries to get data dictionary information. These would be recursive calls. Space management, security checks, calling PL/SQL from SQL—all incur recursive SQL calls.

db block gets：Number of times a CURRENT block was requested.
consistent gets：Number of times a consistent read was requested for a block.
physical reads：Total number of data blocks read from disk. This number equals the value of "physical reads direct" plus all reads into buffer cache.
---------------------------------------------
針對以上3個概念進行的說明解釋及關係如下：
1、DB Block Gets（當前請求的塊數目）
當前模式塊意思就是在操作中正好提取的塊數目，而不是在一致性讀的情況下而產生的塊數。正常的情況下，一個查詢提取的塊是在查詢開始的那個時間點上存在的資料塊，當前塊是在這個時刻存在的資料塊，而不是在這個時間點之前或者之後的資料塊數目。

2、Consistent Gets（資料請求總數在回滾段Buffer中的資料一致性讀所需要的資料塊）
這裡的概念是在處理你這個操作的時候需要在一致性讀狀態上處理多少個塊，這些塊產生的主要原因是因為由於在你查詢的過程中，由於其他會話對資料塊進行操作，而對所要查詢的塊有了修改，但是由於我們的查詢是在這些修改之前呼叫的，所以需要對回滾段中的資料塊的前映像進行查詢，以保證資料的一致性。這樣就產生了一致性讀。

3、Physical Reads（物理讀）
就是從磁碟上讀取資料塊的數量，其產生的主要原因是：
1、 在資料庫快取記憶體中不存在這些塊
2、 全表掃描
3、 磁碟排序

它們三者之間的關係大致可概括為：
邏輯讀指的是Oracle從記憶體讀到的資料塊數量。一般來說是'consistent gets' + 'db block gets'。當在記憶體中找不到所需的資料塊的話就需要從磁碟中獲取，於是就產生了'phsical reads'。

覺得還是不夠詳細，再補充一下：
db block gets :
number of data blocks read in CURRENT mode ie) not in a read consistent fashion, but the current version of the data blocks. 產生db_block_gets的主要方式：
1.DML like Update, Delete will need to access the blocks in the current mode for modification.
2.Most data dictionary calls are done in CURRENT mode.
3.Buffers are often retrieved in current mode for INSERT, UPDATE, and DELETE, and SELECT FOR UPDATE statements. Blocks must be requested in current mode in order to be changed.
4.Also certain classes of blocks, like segment header blocks are always requested in current mode.
5. this will happen when you're doing a FULL TABLE SCAN on a table or a FAST FULL SCAN on an index. In those cases, the segment header is read (usually multiple times for some unknown reason) in current mode.

在current mode模式下產生的對block的訪問叫db block gets，這些block在SGA中，不需要訪問硬碟。
db_block_gets counts logical reads in CURRENT mode in the buffer cache and also includes direct read blocks (sort-on-disk) blocks.
.
consistent gets :
number of data blocks accessed in READ CONSISTENT mode. When a block is requested in consistent mode, if any changes have been committed or made to that block since the requesting statement (or transaction) began, then they must be rolled back for the purposes of the read, to give a consistent view of the data at that time.(當資料庫是在consistent mode模式下被請求的，則假如該資料塊的資料修改是在查詢語句發出之後被提交的，則此查詢語句讀到該資料塊時，還必須為本次查詢將資料回滾，得到查詢語句發出時該資料塊的值，以便給查詢給出一個與該查詢有關的所有資料塊的一致性檢視，這也就是oracle所說的一致性讀) In order to maintain statement level read consistency, Oracle has to read the blocks in a consistent fashion(as of the snapshot SCN) and hence may fetch from rollback segments , which is also added to this statistic. Buffers are usually retrieved in consistent mode for queries. 如普通的select語句、索引訪問而引起的將資料讀入到buffer中(也可能為physical read)或直接從buffer中讀資料。注意，DML語句也能引起consistent gets，如update tab1 set col_b='Hello world' where col_a=1;，因為該語句需要找到需要被修改的所有資料塊，在找資料塊的過程中就會引起consistent gets。
在Read consistent mode模式下產生的對block的訪問叫consistent gets。

Session logical read is: The sum of "db block gets" plus "consistent gets".

db_block_changes:
db_block_changes counts modifications made to CURRENT blocks
'db block changes' = changes made to current blocks under LOGING ( UPDATE, INSERT, DELETE) + changes made to SORT blocks ( NOLOGING).

'consistent changes' :
changes made to block for CONSISTENT READ and changes made to SORT blocks

physical reads :
Physical( disk and/or filesystem page cache) reads. Basically those that cannot be satisfied by the cache and those that are direct reads.
Total number of data blocks read from disk. This number equals the value of "physical reads direct" (direct from disk, excludes buffer cache) plus all reads into buffer cache.

physical writes ：
Total number of data blocks written to disk. This number equals the value of "physical writes direct" (Number of writes directly to disk, bypassing the buffer cache as in a direct load operation) plus all writes from buffer cache.

要注意Oracle 的physical read 與 physical writes並不總是等於硬碟真正意義上的物理讀與物理寫，因為現在都存在作業系統快取記憶體與磁碟子系統快取記憶體，這樣及時I/O沒有被實際寫入磁碟，作業系統I/O子系統或磁碟系統也會確認為一個成功的I/O，所以ORACLE 的physical read 與 physical writes並不是物理上發生讀寫的次數。

redo block size：
redo block size is platform specific. There is a method to determine the size by dumping the redo header, refer to note 154864.1. Redo blocks written does not include archive writes or multiplexed writes.

redo entries:
The redo log buffer is a circular buffer in the SGA that holds information about changes made to the database. This information is stored in redo entries. Redo entries contain the information necessary to reconstruct, or redo, changes made to the database . Redo entries are used for database recovery, if necessary.
Redo entries are copied by Oracle server processes from the user's memory space to the redo log buffer in the SGA. (this copy is what the statistic represents).

db block changes包含dml修改的block，也包含為實現rollback功能而修改的block，但是不包含為實現redo功能而修改的block。

the number of blocks visited = consistent gets + db block gets
the number of blocks visited相當與logical read，即從記憶體中都資料塊的次數。

做一個測試：
用下面語句取得統計資訊：
select sid,value,name from v$sesstat x,v$statname y
where x.statistic#=y.statistic#
and ( name like '%db block%gets%' or name like '%consistent%gets%'
or name like '%consistent%change%' or name like '%db block%change%'
or name like '%physical writes%' or name like '%physical reads%'
or name = 'index fetch by key' or name like '%read%' or name like '%scans%'
)
and sid = 12 and value > 0
order by value;

Lets take an example to illustrate relationship between 'db block gets' and 'db block changes' .

create table toto( i number);

select sid,value,name from v$sesstat x,v$statname y
where x.statistic#=y.statistic#
and ( name like '%db block%gets%' or name like '%consistent%gets%'
or name like '%consistent%change%' or name like '%db block%change%'
or name like '%physical writes%' or name like '%physical reads%'
or name = 'index fetch by key' or name like '%read%' or name like '%scans%'
)
and sid = 12 and value > 0
order by value;

SID VALUE NAME
---------- ---------- ------------------------------
11 95 db block gets
11 108 db block changes
11 0 consistent changes

select count(*) from toto x, toto y, toto z;
COUNT(*)
----------
0

select sid,value,name from v$sesstat x,v$statname y
where x.statistic#=y.statistic#
and ( name like '%db block%gets%' or name like '%consistent%gets%'
or name like '%consistent%change%' or name like '%db block%change%'
or name like '%physical writes%' or name like '%physical reads%'
or name = 'index fetch by key' or name like '%read%' or name like '%scans%'
)
and sid = 11 and value > 0
order by value

SID VALUE NAME
---------- ---------- ------------------------------
11 98 db block gets
11 108 db block changes
11 0 consistent changes

Here we can see that a select increment 'db block gets' by 3 and 'db block changes' remains the same : this illustrates that we can get block in current mode ( to have most recent information) without performing any changes." – 這也可以作為current mode的定義。

下面這段寫的非常好，如果上面的還是不理解，則需要仔細的看一下！
A 'consistent get' is your server process telling the database "I need this dba (data block address) consistent with the point in time represented by this SCN, x."

So, lots of things can happen here. First, Oracle will look in the buffer cache for a CR (consistent read) buffer of the block that's consistent w/ the requested SCN. It may find it, if it does, that's counted as a 'consistent get' and either a 'consistent gets - no work' or 'consistent gets - cleanouts only', depending on whether the block needed to be cleaned out. (See V$SYSSTAT/V$SESSTAT for the statistics.) If it doesn't, it may take an existing CR buffer and roll it back further, or it may clone the current block and roll it back. If it needs to apply rollback (aka undo) then it will increment 'consistent gets' and either 'consistent gets - rollbacks only' or 'consistent gets - cleanouts and rollbacks'.

So, each 'consistent get' is your server process successfully getting access to the contents of a dba consistent w/ a particular SCN. This number should represent the number of buffer gets required to satisfy a particular query.

Now, 'db block gets'. A 'db block get' is a copy of the 'current mode block'. That is, the data in the block, as it exists currently, or at this point in time. Note that while multiple CR copies of a block may exist in the buffer cache, there can only ever be one current mode copy of a block in the buffer cache at any one time. (RAC is a special case, w/ shared current and exclusive current, but I'm not going to get into that here.) So, a 'db block get' is a buffer get in current mode. 'db block gets' are usually associated w/ DML, and in that scenario, will implicitly lock one or more rows in that block. Also, there is a notable case where db block gets can occur with a select statement. That will happen when you're doing a FULL TABLE SCAN on a table or a FAST FULL SCAN on an index. In those cases, the segment header is read (usually multiple times for some unknown reason) in current mode.

Next, 'physical reads': A physical read will occur any time a consistent get or a db block get goes looking for block and can't find it in the buffer cache. So, for each block read from disk, physical reads will be incremented. Gets which result in physical reads are counted both as as get and as a read in the statistics. So, if you do 10 consistent gets and 5 of them require physical reads, you should see consistent gets incremented by 10 and physical reads incremented by 5.

Now, what's up w/ arraysize? Well, arraysize is the client side setting for SQL*Plus specifying the size of the array that will receive result sets. The default, as you learned, is 15. Now, suppose you have a table where there are 30 records per block, and 3,000 rows in the table. But, your arraysize is 15. So, your server process will get the first 15 rows, and return them from the first block buffer. Now, for the next 15, you need to get that same block again, for rows 16-30. So, you need to do two buffer gets per block to get all the rows. For a 3,000 row table, you'll do (approximately) 3,000/15 = 200 buffer gets. If you change your arraysize to 30, you can get away w/ visitng each block only once and do 3,000/30 = 100 buffer gets.

So, consider that even after you've optimized a particular SQL statement, if the arraysize is too small, you're going to force your server process to do excess database calls (FETCH calls), and extra buffer gets as well. This can best be illustrated with a test similar to what you did, but try looking at the raw trace file for FETCH calls. The number of FETCH calls ought to be very close to (number of rows returned / arraysize). The 'r=xxx' in the FETCH call data in the trace file is the number of rows returned, which is probably what your arraysize is set to.

So, db block gets, consistent gets, and physical reads are all measured in buffers (or blocks). If the same block is requested multiple times, it will be counted that many times in these statistics. Oracle will always access data from the buffer cache by the buffer. All 'get' operations are by the buffer, never by row. After the buffer is 'gotten', Oracle parses the block to get the data for the relevant rows.

When a "consistent get" is done, this doesn't necessarily mean that Oracle had to do any rollback/undo of DML in order to get the consistent image. Rather, it just means that Oracle requested a copy of the block as of the point in time that the query was started (a "consistent snapshot"). So, I believe it is quite normal to have many, many consistent gets even if there is NO DML occuring.

自己做的一個更詳細的測試：
用到的查詢統計資訊的語句;
select sid,value,name from v$sesstat x,v$statname y
where x.statistic#=y.statistic#
and ( name like '%db block%gets%' or name like '%consistent%gets%'
or name like '%consistent%change%' or name like '%db block%change%'
or name like '%physical writes%' or name like '%physical reads%'
or name = 'index fetch by key' or name like '%read%' or name like '%scans%'
)
and sid = 11 and value > 0
order by value

Scott使用者登入，但是沒有做任何操作，sid為10，在另外一個使用者下查詢：
SID VALUE NAME
---------- ---------- --------------------------------
11 3 db block gets
11 0 physical reads
11 4 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)

在scott sid=11下執行select * from emp後，結果沒有任何變化，這是不應該的，應該有physical reads與db block gets的值變化

在scott sid = 11下執行update emp set comm = 10; 在另外一個使用者下查詢：
SID VALUE NAME
--------- ---------- ------------------------------
11 18 db block gets
11 0 physical reads
11 33 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)

再次在scott sid = 11的同一個session下執行update emp set comm = 10; 在另外一個使用者下查詢：
SID VALUE NAME
------- ---------- -------------------------------
11 32 db block gets
11 0 physical reads
11 61 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)
不要commit;

用scott再次登入, sid 為17，在另外一個使用者下查詢：
SID VALUE NAME
------- ---------- --------------------------------
17 3 db block gets
17 0 physical reads
17 4 db block changes
17 0 consistent changes
17 0 physical writes
17 0 physical writes non checkpoint
17 0 physical reads direct
17 0 physical writes direct
17 0 physical reads direct (lob)
17 0 physical writes direct (lob)

在scott session 17，下執行select * from emp 後，在另外一個使用者下查詢：
SID VALUE NAME
-------- ---------- --------------------------------
17 3 db block gets
17 0 physical reads
17 5 db block changes
17 28 consistent changes
17 0 physical writes
17 0 physical writes non checkpoint
17 0 physical reads direct
17 0 physical writes direct
17 0 physical reads direct (lob)
17 0 physical writes direct (lob)

再次在scott session 17，下執行select * from emp 後，在另外一個使用者下查詢：
SID VALUE NAME
----- ---------- ---------------------------------
17 3 db block gets
17 0 physical reads
17 6 db block changes
17 56 consistent changes
17 0 physical writes
17 0 physical writes non checkpoint
17 0 physical reads direct
17 0 physical writes direct
17 0 physical reads direct (lob)
17 0 physical writes direct (lob)

在此在scott sid = 11的同一個session下執行rollback; 在另外一個使用者下查詢：
SID VALUE NAME
------ ---------- ------------------------------
11 89 db block gets
11 0 physical reads
11 118 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)

再次在scott sid = 11的同一個session下執行update emp set comm = 10; 在另外一個使用者下查詢：
SID VALUE NAME
---- ---------- --------------------------------
11 104 db block gets
11 0 physical reads
11 147 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)

再次在scott sid = 11的同一個session下執行commit; 在另外一個使用者下查詢：
SID VALUE NAME
---- ---------- ------------------------------
11 105 db block gets
11 0 physical reads
11 148 db block changes
11 0 consistent changes
11 0 physical writes
11 0 physical writes non checkpoint
11 0 physical reads direct
11 0 physical writes direct
11 0 physical reads direct (lob)
11 0 physical writes direct (lob)

經過初步試驗，發現如果開始執行select * from emp後，然後再執行select * from emp where empno = 7902, 不管執行多少此，都不會引起physical reads與db block gets的增加。因為db block gets是在current mode模式中產生的。
而是隻會引起session logical reads與index fetch by key、consistent gets的增加。
如執行
select sid,value,name from v$sesstat x,v$statname y
where x.statistic#=y.statistic#
and ( name like '%db block%change%' or name like '%db block%get%'
or name like '%consistent%change%' or name like '%physical reads%'
or name like '%physical writes%' or name like '%scans%'
or name = 'index fetch by key' or name like '%read%'
)
and sid = 11 and value > 0
order by value
查詢後，得到：
SID VALUE NAME
---- ---------- -------------------------------
11 3 db block gets
11 4 db block changes
11 4 index scans kdiixs1
11 5 index fetch by key
11 6 table scans (short tables)
11 8 no work - consistent read gets
11 42 session logical reads

將資料庫關閉，重新啟動後，重新用scott聯接，sid = 12立即查詢統計資訊，得到：
SID VALUE NAME
---- ---------- -------------------------------
12 6 db block gets
12 6 table scans (short tables)
12 8 db block changes
12 53 cluster key scans
12 89 index fetch by key
12 145 physical reads
12 163 index scans kdiixs1
12 337 consistent gets - examination
12 423 no work - consistent read gets
12 945 consistent gets
12 951 session logical reads

在sid =12 session中執行select * from emp where empno = 7902;後，得到統計資訊：
SID VALUE NAME
---- ---------- -------------------------------
12 6 db block gets
12 6 table scans (short tables)
12 8 db block changes
12 60 cluster key scans
12 98 index fetch by key
12 150 physical reads
12 168 index scans kdiixs1
12 360 consistent gets - examination
12 439 no work - consistent read gets
12 990 consistent gets
12 996 session logical reads
再次執行select * from emp;則db block gets與physical reads不發生變化

將資料庫關閉，重新啟動後，重新用scott聯接，sid = 12立即查詢統計資訊，得到：
SID VALUE NAME
---- ---------- --------------------------------
12 6 db block gets
12 6 table scans (short tables)
12 8 db block changes
12 53 cluster key scans
12 89 index fetch by key
12 145 physical reads
12 163 index scans kdiixs1
12 337 consistent gets - examination
12 423 no work - consistent read gets
12 945 consistent gets
12 951 session logical reads

在sid =12 session中執行select * from emp;後，得到統計資訊：
SID VALUE NAME
--- ---------- -------------------------------
12 6 db block gets
12 7 table scans (short tables)
12 8 db block changes
12 60 cluster key scans
12 97 index fetch by key
12 149 physical reads
12 168 index scans kdiixs1
12 358 consistent gets - examination
12 441 no work - consistent read gets
12 992 consistent gets
12 998 session logical reads
再次執行select * from emp;則db block gets與physical reads不發生變化

轉自：http://blog.csdn.net/rabbitbug/archive/2009/12/16/5015419.aspx

[@more@]