I believe it's my academic background that has seen me develop the need to see the facts for myself before I feel confident about a particular subject. It has always been both a strength and a weakness of mine. On one hand it provides the motivation for me to tackle a subject I want to understand, but on the other I have been guilty of reinventing the wheel just so I can see it work for myself. You might say this is natural for guys like us and essential for personal development however I'm sure my past superiors have seen it unnecessary for me to totally rebuild a system from scratch when all that was required was a simple configuration change. But then where does all the fun come from?!
With this in mind I have finally decided to tackle the subject on fragmentation levels within a database. I've always been told that proper index maintenance is essential for optimal performance, and over the years have employed different methods of achieving this. However I have never really seen proof that my maintenance jobs actually have any benefit. Sure I've seen the odd demonstration and even made improvements to systems I've managed myself but these have been isolated incidents where the customer has made a complaint or a developer didn't take into account the amount of data growth within a particular table. These problems have usually been resolved by reindexing a particular index more frequently or applying a more appropriate index.
What I'm talking about here is a system that has been in production for a number of years, where data volumes have continually grown, application patches applied, OS and SQL service packs and patches applied, etc. The system is considerably different now to what is was when it was first released, and chances are you've taken over the responsibility of these systems somewhere in between.
Ideally a performance audit would need to be carried out periodically to ensure your systems are running optimally, and I imagine this would be quite some task to undertake on some of the systems that are out there.
What I am interested in here is looking at what fragmentation is, how it happens and how we go about resolving it. I am also interested in running a few experiments to see how a SELECT query performs on tables that have indexes with differing fragmentation levels.
This article is more of a discovery voyage rather than a 'how to implement your index maintenance' document and will hopefully give us all a basic foundation to go away with and think about. You might read the entire article and think I could have got this point over in much less text but I want to take you through my entire thought process to aid a better understanding.
After some thought I decided that it was best to revisit the topic on how SQL stores it's data and look at the behavior as data is added. And by running a few experiments we can look further into how the system functions under the surface. Then by adding an index or two we can look at the differences in the behavior that SQL exhibits.
As I said earlier I need to see things for myself and so I am going to start at the very beginning and create a blank database with one table.
-- This script creates our test database and creates a table within it -- You will have to modify the 2 filename parameters to values of where you want the DB created CREATE DATABASE [myIndexes] ON (NAME = N'myIndexes_Data', FILENAME = N'C:\myIndexes_Data.MDF' , SIZE = 1000, FILEGROWTH = 10%) LOG ON (NAME = N'myIndexes_Log', FILENAME = N'C:\myIndexes_Log.LDF' , SIZE = 30, FILEGROWTH = 10%) COLLATE Latin1_General_CI_AS GO USE [myIndexes] GO CREATE TABLE [dbo].[myTable] ( [myPK] [uniqueidentifier] NOT NULL , [myID] [bigint] IDENTITY (1, 1) NOT NULL , [Char1] [varchar] (20) COLLATE Latin1_General_CI_AS NOT NULL , [Char2] [char] (200) COLLATE Latin1_General_CI_AS NOT NULL , [Char3] [varchar] (2000) COLLATE Latin1_General_CI_AS NOT NULL , [Num1] [int] NOT NULL , [Num2] [money] NOT NULL , [Date1] [datetime] NOT NULL ) ON [PRIMARY] GO ALTER DATABASE myIndexes SET RECOVERY SIMPLE GO
Have a look at our database (myIndexes) and you'll see that it is initially set to a 1GB data file, 30MB log file and simple recovery. You might not have enough disk space so feel free to reduce this but it is useful for what we'll do later on.
Looking in the table (myTable) you'll see I've created a number of fields with differing data types. Hopefully we'll be able to see if these have any effect on performance later on in the article. You'll also notice that there aren't any indexes configured on the table - also known as a heap. Initially we are just going to keep things simple and take a look at the basics of the file structure and how SQL maintains it.
OK, so how and where is the database and table (and system objects) actually stored on the disk? And is there any fragmentation at this point?
Well the database and all its objects are obviously stored in the myIndexes.mdf in the location you specified when creating the database. But where is [myTable] stored within this .mdf file?
The objects within a database are stored within data pages which are a little over 8k each. When you create a table you define a number of fields that will store your data. The total number of characters that the combined fields make up cannot exceed this 8k limit. If each character takes up 1 byte then the total number of characters you assign to the table (i.e. each row) cannot be more than 8,000 (8,060 to be exact).
For example if you look at the table we have created in our database you will see it is made up of the following fields:
myPK uniqueidentifier 16 myID bigint 8 Char1 varchar 20 Char2 char 200 Char3 varchar 2000 Num1 int 4 Num2 money 8 Date1 datetime 8 ---- total = 2264
You can see that we have used a total 2,264 bytes. Also I have used the term bytes instead of characters - that's the official way field length is measured. Generally a text field will use 1 byte per character but this does not easily convert for a date (8 bytes) or a money field (8 bytes) so it is best that we refer to the length from now on as bytes. Other data types take up different amounts of storage (e.g.. unicode characters use 2 bytes and would therefore limit you to 4,000 characters per row). I want to keep things relatively simple here so you can look them up in BOL if you want to know more.
Note: Certain types can span multiple pages but we won't be covering this here.
OK, so now you see that in general we're limited to a maximum row size of one data page (8k). More often than not however your total row size will not be anywhere near this and so SQL will be able store more than one row within a single data page.
As you keep adding rows to your table SQL will start to add more data pages and populate them accordingly. Now there's one more thing you should know about your data pages and that is SQL groups them into bundles of 8 - and we refer to these as Extents. An extent is a group of 8 data pages, holding up to 64k of data (i.e. 8k x 8 data pages = 64). Why SQL does this is belong the scope of this article but lets just say it is for optimal disk performance.
Right let's work out how to find which data page is storing our table data. Type the following command... Remember that all the commands that we use will be against the [myIndexes] database unless otherwise stated.
SELECT * FROM SYSINDEXES WHERE NAME = 'myTable'
This command searches for the system information about our table object. You will see that one row is returned containing a number of fields. The fields we're interested in are:
first = a pointer to the first data page
dpages = the number of data pages used for our table data
Currently these values are set to zero because no data has been inserted yet - so lets add our first record by running the following command...
INSERT INTO [myTable] VALUES (NEWID(), 'test1', 'test1', 'test1', 123, 321, GETDATE() )
and confirm it is there...
SELECT * FROM [myTable]
Now we have a record in our database lets re-run the command on the system table to find out about where our data is stored...
SELECT * FROM SYSINDEXES WHERE NAME = 'myTable'
and we now see that these vales have been updated. The values you get may differ from mine so bear this in mind when we use them as parameters for the upcoming exercise. My values are...
first = 0x220000000100
dpages = 1
So this is saying that the data for [myTable] starts on data page 0x220000000100 (a hex value) and that a total of one data page has been used so far.
Now we'll use one of SQL's undocumented commands to view the contents of the data page, but before we do we need to decipher the hex number by doing the following (more information can be found here - http://www.sqlservercentral.com/forums/shwmessage.aspx?forumid=5&messageid=291762).
Split the 0x220000000100 number into two halves which will give us the page id part and the file id part. This gives us page id: 0x220000 and file id: 0x000100. Because we're only using one database file (myIndexes.mdf) then we can be sure that the file id is 1. The page id needs to be reversed which will give us 0x000022 (always leave the 0x bit at the front) and now convert this to decimal. I use Google to do this - just type 'convert 0x000022 to decimal' and I get 34. This is page where our data is stored.
SQL's undocumented command DBCC PAGE allows us to view the contents of a page but we must switch on the trace element first...
DBCC TRACEON (3604) DBCC PAGE (myIndexes, 1, 34, 3)
Remember to replace the value 34 with the one you got from your Google result. Try changing this value to see what else is lurking in other data pages but at this point you'll only see system related data.
I'm not going to discuss this command because it is already well documented on the Internet - see http://support.microsoft.com/kb/q83065/ - you won't find it in BOL and Microsoft say they do not support it seen as it hasn't gone through the usual rigorous testing that supported commands have. Or perhaps it has but they wanted to keep it back as part of their own toolkit for support purposes - I know I've been instructed to use it by Microsoft Product Support services (PSS) in the past while investigating database corruption.
Anyway back to the contents of our data page - you should now see your record within the text and in the bottom section. That's how you view the contents of your data pages. And no, scanning these pages will not reveal system passwords!
So far we have created a blank database, a table and inserted one row of data and have seen that SQL stores this in a data page. Now we need to get a decent amount of data in there so we can start to look at how the system grows and also see if there's anything interesting going on.
I have written a script that populates the table with random data - unfortunately this script does take some time to run but can be stopped and started as you see fit. Lets get this running now and then continue talking about the next step.
Before you do run the script please run the following command and store the results - we'll be talking about this next and we need to capture the values before you start adding the records.
DBCC SHOWCONTIG ('myTable')
Store these results and proceed to run this script to start generating our records...
-- This script will insert records into the database USE [myIndexes] DECLARE @Loop INT DECLARE @LoopTemp1 INT DECLARE @LoopTemp2 INT DECLARE @Char1 VARCHAR(20) DECLARE @Char2 VARCHAR(200) DECLARE @Char3 VARCHAR(2000) DECLARE @Num1 INT DECLARE @Num2 MONEY DECLARE @Date1 DATETIME DECLARE @Char INT SET @Loop = 0 WHILE @Loop < 1000 BEGIN -- Generate Char1 SET @LoopTemp1 = 0 SET @LoopTemp2 = 0 SET @Char1 = '' SET @LoopTemp1 = (RAND()*19)+1 -- Randomise the length of the field WHILE @LoopTemp2 < @LoopTemp1 BEGIN SET @Char1= @Char1 + (char(rand()*26 + 65)) -- Turn this into a character SET @LoopTemp2 = @LoopTemp2 + 1 -- Increase loop value END -- Generate Char2 SET @LoopTemp1 = 0 SET @LoopTemp2 = 0 SET @Char2 = '' SET @LoopTemp1 = (RAND()*199)+1 -- Randomise the length of the field WHILE @LoopTemp2 < @LoopTemp1 BEGIN SET @Char2= @Char2 + (char(rand()*26 + 65)) -- Turn this into a character SET @LoopTemp2 = @LoopTemp2 + 1 -- Increase loop value END -- Generate Char3 SET @LoopTemp1 = 0 SET @LoopTemp2 = 0 SET @Char3 = '' SET @LoopTemp1 = (RAND()*1999)+1 -- Randomise the length of the field WHILE @LoopTemp2 < @LoopTemp1 BEGIN SET @Char3= @Char3 + (char(rand()*26 + 65)) -- Turn this into a character SET @LoopTemp2 = @LoopTemp2 + 1 -- Increase loop value END -- Generate Num1 SET @Num1 = 0 SET @Num1 = RAND() * 100 -- Generate random number -- Generate Num2 SET @Num2 = 0 SET @Num2 = RAND() * 100 -- Generate random number -- Generate Date1 SET @Date1= GETDATE() + (RAND() * 100) -- Turn this into a character INSERT INTO [myTable] VALUES (NEWID(), @Char1, @Char2, @Char3, @Num1, @Num2, @Date1 ) SET @Loop = @Loop + 1 END
You will see at the start of the script that it loops 1000 times (WHILE @Loop < 1000) - this also refers to the number of records that will be added so ideally we want to change this to nearer 500000. Perhaps you can run it overnight. I've challenged a colleague at work to see which of us can come up with a script that randomly inserts values into the database the fastest - I have some good ideas but not had the time to try them out. If anyone is interested in joining in the fun (or has any suggestions) let me know and I'll send you the contest rules!
OK while the script is running lets run a check to make sure we can see things happening as we would expect them to. Run the following to check that the number of data pages has been increasing - remember it was at one before...
SELECT * FROM SYSINDEXES WHERE NAME = 'myTable'
If you repeatedly run this the number of rowcnt and dpages should keep going up. However depending on how busy your system is you may need to issue the following command to force a system update. Remember this command because it is very useful when you need up to date information.
If you've ever resized your database or inserted a number of records and SQL does not report values to represent this then running this command will update SQL's internal system values for you. We might see more of this in action later.
BOL definition: Reports and corrects inaccuracies in the sysindexes table, which may result in incorrect space usage.
If you remember earlier on one of the questions we asked is whether we had any fragmentation after the point we had inserted a record. Well I'm sure you'd agree that this is highly unlikely since we are dealing with a brand new database. And before you ran the script that inserts lots of records I asked that you quickly run the DBCC SHOWCONTIG command and store the results. Well now we'll look at what we actually did here...
DBCC SHOWCONTIG, as BOL puts it, displays fragmentation information for the data and indexes of the specified table. We actually ran this command against our table and would have received results like the following:
- Pages Scanned................................: 1 - Extents Scanned..............................: 1 - Extent Switches..............................: 0 - Avg. Pages per Extent........................: 1.0 - Scan Density [Best Count:Actual Count].......: 100.00% [1:1] - Extent Scan Fragmentation ...................: 0.00% - Avg. Bytes Free per Page.....................: 7827.0 - Avg. Page Density (full).....................: 3.30%
Lets look at each of the results in turn to gain an understanding of what it is they are telling us.
- Pages Scanned
This tells us that 1 page was scanned in total to produce our results. And we are happy with the fact that we only have 1 page.
As we know there are 8 pages per extent but seen as we only have 1 page then that would infer we only have 1 extent.
If there were 100 extents SQL might have to flick between each of them multiple times to read all the records in the data pages. When performing this command think of it that SQL must access every record therefore having to visit every page and extent that contain current records. If the records are not in order then SQL will need to keep flicking between Extents to access the required data. You see when a record is added it isn't necessarily added to the end of the table - it might be inserted somewhere in the middle. Or if a record is updated and it becomes too big to fit in to it's current space allocation SQL will move it and set a pointer to the new location. Therefore when it comes to reading it again it will be passed to the new location which may be on another page and extent. It is this flicking that causes additional disk reads which in turn affects performance.
Avg. Pages per Extent
As we know we only have one data page so SQL has only create one at this point. As data starts to grow SQL will think ahead and create more.
Best count is the ideal number of extent changes if everything is contiguously linked. Actual count is the actual number of extent changes. The number in scan density is 100 if everything is contiguous; if it is less than 100, some fragmentation exists. Scan density is a percentage. As you can see from our results we have a perfect 100%!
Extent Scan Fragmentation
Because we're not using an index yet this value is irrelevant.
Avg. Bytes Free per Page
Average number of free bytes on the pages scanned. The higher the number, the less full the pages are. Lower numbers are better.
Avg. Page Density (full)
Average page density (as a percentage). This value takes into account row size, so it is a more accurate indication of how full your pages are. The higher the percentage, the better.
Ideally the number of extents scanned should be one more than the number of extent switches that take place. This would mean that SQL has started at extent one and read through each of them in turn without being redirected elsewhere.
If you run the DBCC SHOWCONTIG command again on your table you will start to see different results on the extents. Because we are working with a heap (a table without a clustered index) the data is in no particular order. It is generally in the order that it was inserted into the table (providing that we are mainly inserting new rows) however SQL will also insert data in to any available space it comes across. If you do a SELECT * FROM [myTable] you will probably find that the [myID] column doesn't increment serially in the early part of the table and this is because as SQL grows the pages and extents it will copy approximately half of the contents of one page into the next so that there's always room for quick insertion of additional records later on (known as page splits). In the early stages of our data generation script SQL will find free space in some of the initially created data pages and insert some records into these pages rather than at then end of the table. I've never actually had confirmation of this but this is how I understand it - it's as if SQL initially expands quite quickly leaving a lot of free space in the first extent(s) and then returns shortly after to fill them in.
I hope you have picked up on the point that because we are using a heap the data is in no particular order. And if you perform any kind of SELECT statement on a heap the entire table must be scanned. And as we have read if we were to select all the records from the table we wouldn't necessarily get them back in the order they were inserted.
Saying that, and I hope you are with me at this point, can a heap actually become fragmented? And if so is there a command to defrag it?
Well I suppose the answer to whether a heap can become fragmented is yes but would that actually cause us any problems? It's not as if we would be scanning the table without any indexes applied and, as we've not specified any order for the table, then we're not loosing out. Fragmentation, of a kind, can occur on a heap when forwarding records are created that point to an updated record that was moved due to the update making it bigger that it's currently allocated space (i.e. there was no room for it in the current slot).
There is no command that will defrag a heap (unless you apply a clustered index and remove it again as this will force an order which will remain after the index is dropped, but will not be maintained).
So we've not really learnt anything new about how to improve our table fragmentation but we have been through the motions of how data builds up within the database file. And we've seen that having a table without an index, in particular a clustered index, deserves the name 'heap' because basically that is all it is, and of little use to us. Running any form of query on a heap will force a full scan of the table and could take a considerable amount of time. And the results from the DBCC SHOWCONTIG command, which is a tool for measuring fragmentation, is also of little use on a heap.
I am going to finish at this point because it is a big subject to tackle and I want to get some feedback before releasing the next part of this article. In my next installment we will look further into data pages and start analyzing the different types and how these manage the allocation of extents and free space. Then we'll start using indexes. We're a long way off but we'll get there slowly and in an academic fashion.
Do have a play and a think about what we have covered so far. Any suggestions would also be welcome. And please remember I am by no means a SQL expert - this work is my findings that I am willing to share with you. If you feel I have mislead you, have incorrectly stated something or simply need more info then please submit feedback.
Until next time...