Improve SQL Server Performance with Compression

RSS

Improve SQL Server Performance with Compression

24 Nov

SQL Server data compression is now available in all editions of SQL Server, starting with 2016 SP1.

In this Article, you will not only know how data compression will save space, you’ll also find out how compression can sometimes improve performance as well.

Space Savings vs. Performance

When I first heard about compression back in 2008, my first thought is that it would have a big performance penalty. Compression would save disk space, but it would probably decrease performance as the data was compressed and decompressed. It turns out that compression can improve performance instead. Because compressed data fits in a smaller number of data pages, there are decreased I/O requirements. Since I/O is generally the bottleneck in SQL Server, this can improve performance. Compressed data also has a decreased memory requirement. When querying compressed data, a smaller number of pages will be copied to the buffer pool. The one area that is impacted is CPU. You do need to have some CPU headroom because compression will require some additional CPU resources.

The good thing is that if the workload is reasonably tuned, many SQL Server instances have more CPU resources than they need. One note of caution here. Don’t look to compression as the solution to solving major performance issues. You need to look at physical resources, configuration, indexing, and query tuning. The point I’m trying to make is that decreased I/O and better memory utilization will benefit the workload in many cases.

Availability

Compression has been available in SQL Server for all versions from 2008 to 2016 SQL server, but only for Enterprise Edition. Beginning with SQL Server 2016 SP1, it’s now available in Enterprise, Standard, and Express. I’m excited about this because now compression is within the reach of any organization.

Compression Types

There are two types of compression that you can use to compress a table, index, or even a partition:

Row Level:
Row level compression works by storing fixed-width data types as variable length data types. Nulls and zeroes do not take any space. For example, in an uncompressed table, an integer column takes 4 bytes per row, even for those rows that have a small number such as 1 or 10, or even null. Once row compression is implemented on the table, each value will be stored with the smallest possible number of bytes. So outside of the metadata, storing a 1 in an integer column will take 1 byte, storing a null or 0 will take 0 bytes.

Starting with SQL Server 2012, Unicode compression is implemented when you use row compression. This applies to Unicode columns in varchar and in char. In an uncompressed table, each character takes up 2 bytes in a Unicode column, even if a small character set is used. In those cases, compressing the table will store the Unicode characters in 1 byte instead.

In the technical article, Data Compression: Strategy, Capacity Planning, and Best Practices, Microsoft recommends using row compression on all data, as long as you have 10% extra CPU capacity, and, of course, as long as you achieve some space savings.
It costs Low CPU penalty.
Page Level:
When you implement page level compression, the rows are automatically row-compressed first.
Page level compression also removes repeated data within a page by two mechanisms, prefix and dictionary compression.
Page level compression can compress a table to a smaller size than row compression, but it is recommended for tables that are mostly inserted, but not updated that often.
It does have a higher CPU penalty.

Neither of these compression types will work on row-overflow data. This is data from a row that exceeds 8K. For example, you can create a table that has two varchar 8000 columns. If a row exceeds 8060 bytes, then SQL Server will move one or more of those columns to another page so that the row fits. One way to get around this issue is the new COMPRESS function.

COMPRESS Function

You can use this function to compress individual values, which includes row-overflow data.
It uses a GZIP algorithm to compress the values. There is a downside to this, however. The COMPRESS function must be applied each time a value is inserted or updated, and a DECOMPRESS function must then be used to read the value. This means that there would be changes to the application or stored procedure.

Performance Metrics

In this demo, we’ll do some performance comparisons. We’ll take a close look at I/O and memory impact. We’ll see how compression affects both reads and writes to the data.

Select comparison

I created two tables; one with page compression called “bigTransactionHistoryPAGE” and one with row compression called “bigTransactionHistoryROW”. I also created a table with no compression called “bigTransactionHistory” that I’ll use during this demonstration as well. And finally, I populated the two compressed tables with around 31 million records.

Let’s take a look at the number of pages in each table or index:

SELECT OBJECT_NAME(i.[object_id]) AS TableName,

       i.name AS IndexName, SUM(s.used_page_count) IndexPages,

       FORMAT(1 - SUM(s.used_page_count) * 1.0/CASE WHEN i.name LIKE 'IX%' THEN 131819 ELSE 143645 END, 'P') AS PercentSaved

FROM sys.dm_db_partition_stats  AS s

JOIN sys.indexes AS i

ON s.[object_id] = i.[object_id] AND s.index_id = i.index_id

WHERE OBJECT_NAME(i.[object_id]) LIKE 'bigTransactionHistory%'

       AND OBJECT_NAME(i.[object_id]) <> 'bigTransactionHistoryTEST'

GROUP BY i.[object_id],i.name

ORDER BY IndexPages Desc;

The uncompressed table from is the largest. The non-clustered index compressed with row compression saved about 30% of the space, and the clustered index saved about 37%. The page compressed table is really interesting. The clustered index is much smaller than the original table, while the non-clustered index is about 39% smaller.

I’m going to turn on STATISTICS IO to compare the number of pages touched when I query each table:

SET STATISTICS IO ON;

GO

SELECT SUM(Quantity) AS ItemsPurchased

FROM bigTransactionHistory;




SELECT SUM(Quantity) AS ItemsPurchased

FROM bigTransactionHistoryROW;




SELECT SUM(Quantity) AS ItemsPurchased

FROM bigTransactionHistoryPAGE;

You see here that less I/O is needed for the compressed table. That’s actually obvious since the clustered and non-clustered indexes are smaller when compressed.

The next query shows how many pages are in the buffer for each index:

SELECT COUNT(*)AS cached_pages_count

,name ,index_id

FROM sys.dm_os_buffer_descriptors AS bd

INNER JOIN

(

SELECT object_name(object_id) AS name

,index_id ,allocation_unit_id

FROM sys.allocation_units AS au

INNER JOIN sys.partitions AS p

ON au.container_id = p.hobt_id

AND (au.type = 1 OR au.type = 3)

UNION ALL

SELECT object_name(object_id) AS name

,index_id, allocation_unit_id

FROM sys.allocation_units AS au

INNER JOIN sys.partitions AS p

ON au.container_id = p.partition_id

AND au.type = 2

) AS obj

ON bd.allocation_unit_id = obj.allocation_unit_id

WHERE database_id = DB_ID()  AND name LIKE 'bigTransactionHistory%'

GROUP BY name, index_id

ORDER BY cached_pages_count DESC;

Again, when the index resides on fewer pages, the data takes less space and memory. Notice that the clustered index was used for the page-compressed table. We saw that the page-compressed clustered index is much smaller than the non-clustered index.

On this Azure VM, I have seen inconsistent results in the time to run the queries. For example, if I use DBCC DROPCLEANBUFFERS, or even restart SQL Server to ensure the data must be loaded from disk, one of these queries could run anywhere from 1 second to even a minute. I suspect that the VM is sharing resources with other VMs, and that is causing my discrepancies. I decided to run the queries each in a loop to better see the difference. Inside the loop, I use DBCC DROPCLEANBUFFERS to make sure that the data was loaded from disk, and not cached data.

DECLARE @Count INT = 0;

WHILE @Count < 100 BEGIN

       DBCC DROPCLEANBUFFERS;    

       SELECT SUM(Quantity) AS ItemsPurchased

       FROM bigTransactionHistory;

       SET @Count += 1;

END;

The loop against the uncompressed table took 3 minutes and 31 seconds. The row-compressed table took 3 minutes and 4 seconds and the page compressed table took 3 minutes and 18 seconds. So in this case, the row-compressed query performed the best, with the page-compressed query next.

Insert comparison

Let’s take a look at inserting data. In this script, I populated an uncompressed table and our two compressed tables with a loop of 1000 inserts of 1000 rows each. To make things fair, I ran DROPCLEANBUFFERS before each loop.

DBCC DROPCLEANBUFFERS;

GO

--2:40

DECLARE @Count INT = 0;

WHILE @Count < 1000 BEGIN

INSERT INTO dbo.bigTransactionHistoryTEST

SELECT *

FROM bigTransactionHistory

WHERE TransactionID BETWEEN @Count * 1000 + 1 AND (@Count + 1) * 1000;

SET @Count = @Count + 1;

END;

The uncompressed inserts took 2 minutes and 40 seconds. The row-compressed inserts completed a bit faster at 2 minutes and 23 seconds, and the page-compressed inserts completed in 2 minutes and 14 seconds. The page-compressed inserts were actually fastest, and I suspect that it might be due to less I/O.

Update comparison

Let’s take a look at updates.

DBCC DROPCLEANBUFFERS;

–1:48

DECLARE @Count INT = 0;

WHILE @Count < 1000 BEGIN

UPDATE dbo.bigTransactionHistoryTEST

SET Quantity = Quantity + 1

WHERE TransactionID BETWEEN @Count * 1000 + 1 AND (@Count + 1) * 1000;

SET @Count = @Count + 1;

END;

Again, I ran loops. In this case, it’s 1000 updates of 1000 rows Updating the uncompressed table took 1 minute and 48 seconds. Updating the row-compressed table took 1 minute and 36 seconds. The page-compressed table was much slower at 2 minutes and 21 seconds. During the slides, I mentioned that Microsoft recommends page compression for workloads that have few updates, and you can see that the updates are slower for page compression.

To summarize, I saw decent space savings, especially with page compression. The compressed tables performed better for both selects and inserts. Updates, however, performed noticeably worse with page compression.

Summary:

This Article is an overview of data compression. Compression is now available in all editions of SQL Server, starting with 2016 SP1. You can implement row and page level compression, and also use the new COMPRESS function. Compression can improve performance because of decreased I/O and memory pressure. I hope this article has been informative for you.

SQL Server Consultation