SQL Server 2008+聚簇索引的排序顺序

Question

Does the sort order of a SQL Server 2008+ clustered index impact the insert performance?

The datatype in the specific case is integer and the inserted values are ascending ( Identity ). Therefore, the sort order of the index would be opposite to the sort order of the values to be inserted.

My guess is, that it will have an impact, but I don't know, maybe SQL Server has some optimizations for this case or it's internal data storage format is indifferent to this.

Please note that the question is about the INSERT performance, not SELECT .

Update
To be more clear about the question: What happens when the values which will be inserted ( integer ) are in reverse order ( ASC ) to the ordering of the clustered index ( DESC )?

Answer 1

There is a difference. Inserting out of Cluster Order causes massive fragmentation.

When you run the following code the DESC clustered index is generating additional UPDATE operations at the NONLEAF level.

CREATE TABLE dbo.TEST_ASC(ID INT IDENTITY(1,1) 
                            ,RandNo FLOAT
                            );
GO
CREATE CLUSTERED INDEX cidx ON dbo.TEST_ASC(ID ASC);
GO

CREATE TABLE dbo.TEST_DESC(ID INT IDENTITY(1,1) 
                            ,RandNo FLOAT
                            );
GO
CREATE CLUSTERED INDEX cidx ON dbo.TEST_DESC(ID DESC);
GO

INSERT INTO dbo.TEST_ASC VALUES(RAND());
GO 100000

INSERT INTO dbo.TEST_DESC VALUES(RAND());
GO 100000

The two Insert statements produce exactly the same Execution Plan but when looking at the operational stats the differences show up against [nonleaf_update_count].

SELECT 
OBJECT_NAME(object_id)
,* 
FROM sys.dm_db_index_operational_stats(DB_ID(),OBJECT_ID('TEST_ASC'),null,null)
UNION
SELECT 
OBJECT_NAME(object_id)
,* 
FROM sys.dm_db_index_operational_stats(DB_ID(),OBJECT_ID('TEST_DESC'),null,null)

There is an extra –under the hood- operation going on when SQL is working with DESC index that runs against the IDENTITY. This is because the DESC table is becoming fragmented (rows inserted at the start of the page) and additional updates occur to maintain the B-tree structure.

The most noticeable thing about this example is that the DESC Clustered Index becomes over 99% fragmented. This is recreating the same bad behaviour as using a random GUID for a clustered index. The below code demonstrates the fragmentation.

SELECT 
OBJECT_NAME(object_id)
,* 
FROM sys.dm_db_index_physical_stats  (DB_ID(), OBJECT_ID('dbo.TEST_ASC'), NULL, NULL ,NULL) 
UNION
SELECT 
OBJECT_NAME(object_id)
,* 
FROM sys.dm_db_index_physical_stats  (DB_ID(), OBJECT_ID('dbo.TEST_DESC'), NULL, NULL ,NULL) 

UPDATE:

On some test environments I'm also seeing that the DESC table is subject to more WAITS with an increase in [page_io_latch_wait_count] and [page_io_latch_wait_in_ms]

UPDATE:

Some discussion has arisen about what is the point of a Descending Index when SQL can perform Backward Scans. Please read this article about the limitations of Backward Scans .

Answer 2

The order of values inserted into a clustered index most certainly impacts performance of the index, by potentially creating a lot of fragmentation, and also affects the performance of the insert itself.

I've constructed a test-bed to see what happens:

USE tempdb;

CREATE TABLE dbo.TestSort
(
    Sorted INT NOT NULL
        CONSTRAINT PK_TestSort
        PRIMARY KEY CLUSTERED
    , SomeData VARCHAR(2048) NOT NULL
);

INSERT INTO dbo.TestSort (Sorted, SomeData)
VALUES  (1797604285, CRYPT_GEN_RANDOM(1024))
    , (1530768597, CRYPT_GEN_RANDOM(1024))
    , (1274169954, CRYPT_GEN_RANDOM(1024))
    , (-1972758125, CRYPT_GEN_RANDOM(1024))
    , (1768931454, CRYPT_GEN_RANDOM(1024))
    , (-1180422587, CRYPT_GEN_RANDOM(1024))
    , (-1373873804, CRYPT_GEN_RANDOM(1024))
    , (293442810, CRYPT_GEN_RANDOM(1024))
    , (-2126229859, CRYPT_GEN_RANDOM(1024))
    , (715871545, CRYPT_GEN_RANDOM(1024))
    , (-1163940131, CRYPT_GEN_RANDOM(1024))
    , (566332020, CRYPT_GEN_RANDOM(1024))
    , (1880249597, CRYPT_GEN_RANDOM(1024))
    , (-1213257849, CRYPT_GEN_RANDOM(1024))
    , (-155893134, CRYPT_GEN_RANDOM(1024))
    , (976883931, CRYPT_GEN_RANDOM(1024))
    , (-1424958821, CRYPT_GEN_RANDOM(1024))
    , (-279093766, CRYPT_GEN_RANDOM(1024))
    , (-903956376, CRYPT_GEN_RANDOM(1024))
    , (181119720, CRYPT_GEN_RANDOM(1024))
    , (-422397654, CRYPT_GEN_RANDOM(1024))
    , (-560438983, CRYPT_GEN_RANDOM(1024))
    , (968519165, CRYPT_GEN_RANDOM(1024))
    , (1820871210, CRYPT_GEN_RANDOM(1024))
    , (-1348787729, CRYPT_GEN_RANDOM(1024))
    , (-1869809700, CRYPT_GEN_RANDOM(1024))
    , (423340320, CRYPT_GEN_RANDOM(1024))
    , (125852107, CRYPT_GEN_RANDOM(1024))
    , (-1690550622, CRYPT_GEN_RANDOM(1024))
    , (570776311, CRYPT_GEN_RANDOM(1024))
    , (2120766755, CRYPT_GEN_RANDOM(1024))
    , (1123596784, CRYPT_GEN_RANDOM(1024))
    , (496886282, CRYPT_GEN_RANDOM(1024))
    , (-571192016, CRYPT_GEN_RANDOM(1024))
    , (1036877128, CRYPT_GEN_RANDOM(1024))
    , (1518056151, CRYPT_GEN_RANDOM(1024))
    , (1617326587, CRYPT_GEN_RANDOM(1024))
    , (410892484, CRYPT_GEN_RANDOM(1024))
    , (1826927956, CRYPT_GEN_RANDOM(1024))
    , (-1898916773, CRYPT_GEN_RANDOM(1024))
    , (245592851, CRYPT_GEN_RANDOM(1024))
    , (1826773413, CRYPT_GEN_RANDOM(1024))
    , (1451000899, CRYPT_GEN_RANDOM(1024))
    , (1234288293, CRYPT_GEN_RANDOM(1024))
    , (1433618321, CRYPT_GEN_RANDOM(1024))
    , (-1584291587, CRYPT_GEN_RANDOM(1024))
    , (-554159323, CRYPT_GEN_RANDOM(1024))
    , (-1478814392, CRYPT_GEN_RANDOM(1024))
    , (1326124163, CRYPT_GEN_RANDOM(1024))
    , (701812459, CRYPT_GEN_RANDOM(1024));

The first column is the primary key, and as you can see the values are listed in random(ish) order. Listing the values in random order should make SQL Server either:

1. Sort the data, pre-insert
2. Not sort the data, resulting in a fragmented table.

The CRYPT_GEN_RANDOM() function is used to generate 1024 bytes of random data per row, to allow this table to consume multiple pages, which in turn allows us to see the effects of fragmented inserts.

Once you run the above insert, you can check fragmentation like this:

SELECT * 
FROM sys.dm_db_index_physical_stats(DB_ID(), OBJECT_ID('TestSort'), 1, 0, 'SAMPLED') ips;

Running this on my SQL Server 2012 Developer Edition instance shows average fragmentation of 90%, indicating SQL Server did not sort during the insert.

The moral of this particular story is likely to be, "when in doubt, sort, if it will be beneficial". Having said that, adding and ORDER BY clause to an insert statement does not guarantee the inserts will occur in that order. Consider what happens if the insert goes parallel, as an example.

On non-production systems you can use trace flag 2332 as an option on the insert statement to "force" SQL Server to sort the input prior to inserting it. @PaulWhite has an interesting article, Optimizing T-SQL queries that change data covering that, and other details. Be aware, that trace flag is unsupported, and should NOT be used in production systems, since that might void your warranty. In a non-production system, for your own education, you can try adding this to the end of the INSERT statement:

OPTION (QUERYTRACEON 2332);

Once you have that appended to the insert, take a look at the plan, you'll see an explicit sort:

It would be great if Microsoft would make this a supported trace flag.

Paul White made me aware that SQL Server does automatically introduce a sort operator into the plan when it thinks one will be helpful. For the sample query above, if I run the insert with 250 items in the values clause, no sort is implemented automatically. However, at 251 items, SQL Server automatically sorts the values prior to the insert. Why the cutoff is 250/251 rows remains a mystery to me, other than it seems to be hard-coded. If I reduce the size of the data inserted in the SomeData column to just one byte, the cutoff is still 250/251 rows, even though the size of the table in both cases is just a single page. Interestingly, looking at the insert with SET STATISTICS IO, TIME ON; shows the inserts with a single byte SomeData value take twice as long when sorted.

Without the sort (ie 250 rows inserted):

SQL Server parse and compile time: 
   CPU time = 0 ms, elapsed time = 0 ms.
SQL Server parse and compile time: 
   CPU time = 16 ms, elapsed time = 16 ms.
SQL Server parse and compile time: 
   CPU time = 0 ms, elapsed time = 0 ms.
Table 'TestSort'. Scan count 0, logical reads 501, physical reads 0, 
   read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob 
   read-ahead reads 0.

(250 row(s) affected)

(1 row(s) affected)

 SQL Server Execution Times:
   CPU time = 0 ms,  elapsed time = 11 ms.

With the sort (ie 251 rows inserted):

SQL Server parse and compile time: 
   CPU time = 0 ms, elapsed time = 0 ms.
SQL Server parse and compile time: 
   CPU time = 15 ms, elapsed time = 17 ms.
SQL Server parse and compile time: 
   CPU time = 0 ms, elapsed time = 0 ms.
Table 'TestSort'. Scan count 0, logical reads 503, physical reads 0, 
   read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob 
   read-ahead reads 0.
Table 'Worktable'. Scan count 0, logical reads 0, physical reads 0, 
   read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob 
   read-ahead reads 0.

(251 row(s) affected)

(1 row(s) affected)

 SQL Server Execution Times:
   CPU time = 16 ms,  elapsed time = 21 ms.

Once you start to increase the row size, the sorted version certainly becomes more efficient. When inserting 4096 bytes into SomeData , the sorted insert is nearly twice as fast on my test rig as the unsorted insert.

---

As a side-note, in case you're interested, I generated the VALUES (...) clause using this T-SQL:

;WITH s AS (
    SELECT v.Item
    FROM (VALUES (0), (1), (2), (3), (4), (5), (6), (7), (8), (9)) v(Item)
)
, v AS (
    SELECT Num = CONVERT(int, CRYPT_GEN_RANDOM(10), 0)
)
, o AS (
    SELECT v.Num
        , rn = ROW_NUMBER() OVER (PARTITION BY v.Num ORDER BY NEWID())
    FROM s s1
        CROSS JOIN s s2
        CROSS JOIN s s3
        CROSS JOIN v 
)
SELECT TOP(50) ', (' 
        + REPLACE(CONVERT(varchar(11), o.Num), '*', '0') 
        + ', CRYPT_GEN_RANDOM(1024))'
FROM o
WHERE rn = 1
ORDER BY NEWID();

This generates 1,000 random values, selecting only the top 50 rows with unique values in the first column. I copied-and-pasted the output into the INSERT statement above.

Answer 3

As long as the data comes ordered by the clustered index (irrespective if it's ascending or descending), then there should not be any impact on the insert performance. The reasoning behind this is that SQL does not care of the physical order of the rows in a page for the clustered index. The order of the rows is kept in what is called a "Record Offset Array", which is the only one that needs to be rewritten for a new row (which anyway would have been done irrespective of order). The actual data rows will just get written one after the other.

At transaction log level, the entries should be identical irrespective of the direction so this will not generate any additional impact on performance. Usually the transaction log is the one that generates most of the performance issues, but in this case there will be none.

You can find a good explanation on the physical structure of a page / row here https://www.simple-talk.com/sql/database-administration/sql-server-storage-internals-101/ .

So basically as long as your inserts will not generate page splits (and if the data comes in the order of the clustered index irrespective of order it will not), your inserts will have negligible if any impact on the insert performance.

Answer 4

Based on the code below, inserting data into an identity column with a sorted clustered index is more resource intense when the selected data is ordered in the opposite direction of the sorted clustered index.

In this example, logical reads are nearly double.

After 10 runs, the sorted ascending logical reads average 2284 and the sorted descending logical reads average 4301.

--Drop Table Destination;
Create Table Destination (MyId INT IDENTITY(1,1))

Create Clustered Index ClIndex On Destination(MyId ASC)

set identity_insert destination on 
Insert into Destination (MyId)
SELECT TOP (1000) n = ROW_NUMBER() OVER (ORDER BY [object_id]) 
FROM sys.all_objects 
ORDER BY n


set identity_insert destination on 
Insert into Destination (MyId)
SELECT TOP (1000) n = ROW_NUMBER() OVER (ORDER BY [object_id]) 
FROM sys.all_objects 
ORDER BY n desc;

More about logical reads if you are interested: https://www.brentozar.com/archive/2012/06/tsql-measure-performance-improvements/