6

I'm setting up an OLAP database using SQL Server 2014. The core fact table has about 40,000,000 rows, 225 columns with an average row size of 181 bytes. I've been toying around with the clustered columnstore index a bit without much luck. In general I find query performance more than 4 times slower with the new technology.

One particular example - selecting a single row using an int32 primary key now takes 12 seconds ... this is a sub-second operation on a rowstore table (of course it had a unique index on the PK, which is not allowed in conjunction with the clustered columnstore index).

I'm trying to figure out what I'm doing wrong - from the MS docs it sounds like this is ideal technology for this task; maybe I'm missing something.

I'm running SQL 2014 Enterprise on Windows 8.1 64-bit with 128GB RAM and SSD for data storage. The data is read only for this app.

  • 5
    Columnstore indexes are not going to outperform B trees for singleton seeks. Have you compared times to run an aggregation query on the whole table? – Martin Smith Sep 16 '15 at 20:38
  • Yes I did some comparisons there also - typically 3-6 times slower with the columnstore vs. covered queries in rowstore mode. I'm guessing I've got some sort of configuration problem. – Hank Brandenburg Sep 16 '15 at 20:47
  • Columnstore will give you a boost only if your execution mode is batch mode. Can you look into execution plan of the query and see if it is batch mode or row mode ? Also, what is MAXDOP setting on the instance ? – Kin Shah Sep 17 '15 at 1:18
7

If you can post the specific data and queries you are using, that is probably the only way we can help answer the question in the context of your specific case. You can use a script that generates anonymous data in roughly the same scale as your real example.

However, I went ahead and created a similar type of script myself. For the sake of simplicity, I am using fewer than 225 columns. But I am using the same number of rows and random data (which is unfavorable for columnstore) and I saw results that are much different than yours. So my initial thought is that yes, you do have some sort of problem with either your configuration or your test queries.

A few of the key takeaways:

  • Columnstore has dramatically faster performance than rowstore for simple aggregations across all rows in a column
  • If loaded carefully, columnstore can perform surprisingly well for singleton seeks. There is an I/O hit, but with a warm cache performance was very good. But not as good as rowstore for this use case, of course.
  • If you need to be able to perform both singleton seeks and large aggregation queries, you might consider using a non-clustered columnstore index on top of a standard b-tree table.
  • You mention that you have 225 columns, but an average row is just 181 bytes. This seems a little unusual; is your table mostly BIT columns? That might be something to look into further. I did see very good compression ratios on a simple BIT column columnstore (over 99%), but it may be the case that much of that is due to the absence of row overhead and this advantage would disappear with many BIT columns on a single row.
  • If you want to learn (a lot) more about columnstore, Niko's 66-part (and counting) blog series has been the most valuable reference that I've come across.

And now on to the details:

Create rowstore data set

Nothing too exciting here; we create 40MM rows of pseudo-random data.

SELECT @@VERSION
--Microsoft SQL Server 2014 - 12.0.4213.0 (X64) 
--  Jun  9 2015 12:06:16 
--  Copyright (c) Microsoft Corporation
--  Developer Edition (64-bit) on Windows NT 6.1 <X64> (Build 7601: Service Pack 1)
GO

-- Create a rowstore table with 40MM rows of pseudorandom data
;WITH E1(N) AS (
    SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1 
    UNION ALL SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1 UNION ALL SELECT 1
)
, E2(N) AS (SELECT 1 FROM E1 a CROSS JOIN E1 b)
, E4(N) AS (SELECT 1 FROM E2 a CROSS JOIN E2 b)
, E8(N) AS (SELECT 1 FROM E4 a CROSS JOIN E4 b)
SELECT TOP 40000000 ISNULL(ROW_NUMBER() OVER (ORDER BY (SELECT NULL)), 0) AS id
    , ISNULL((ABS(CAST(CAST(NEWID() AS VARBINARY) AS INT)) % 5) + 1, 0) AS col1
    , ISNULL(ABS(CAST(CAST(NEWID() AS VARBINARY) AS INT)) * RAND(), 0) AS col2
    , ISNULL(ABS(CAST(CAST(NEWID() AS VARBINARY) AS INT)) * RAND(), 0) AS col3
    , ISNULL(ABS(CAST(CAST(NEWID() AS VARBINARY) AS INT)) * RAND(), 0) AS col4
    , ISNULL(ABS(CAST(CAST(NEWID() AS VARBINARY) AS INT)) * RAND(), 0) AS col5
INTO dbo.test_row
FROM E8
GO
ALTER TABLE test_row
ADD CONSTRAINT PK_test_row PRIMARY KEY (id)
GO

Create columnstore data set

Let's create the same data set as a CLUSTERED COLUMNSTORE, using the techniques described to load data for better segment elimination on Niko's blog.

-- Create a columnstore table with the same 40MM rows
-- The data is first ordered by id and then a single thread
-- use to build the columnstore for optimal segment elimination
SELECT *
INTO dbo.test_column
FROM dbo.test_row
GO
CREATE CLUSTERED INDEX cs_test_column
ON dbo.test_column (id)
GO
CREATE CLUSTERED COLUMNSTORE INDEX cs_test_column 
ON dbo.test_column WITH (DROP_EXISTING = ON, MAXDOP = 1)
GO

Size comparison

Because we are loading random data, columnstore achieves only a modest reduction in table size. If the data was not as random, the columnstore compression would dramatically decrease the size of the columnstore index. This particular test case is actually quite unfavorable for columnstore, but it's still nice to see that we get a little bit of compression.

-- Check the sizes of the two tables
SELECT t.name, ps.row_count, (ps.reserved_page_count*8.0) / (1024.0) AS sizeMb
FROM sys.tables t WITH (NOLOCK)
JOIN sys.dm_db_partition_stats ps WITH (NOLOCK)
    ON ps.object_id = t.object_id
WHERE t.name IN ('test_row','test_column')
--name          row_count   sizeMb
--test_row      40000000    2060.6328125
--test_column   40000000    1352.2734375
GO

Performance comparison

In the following two test cases, I try two very different use cases.

The first is the singleton seek mentioned in your question. As commenters point out, this is not at all the use case for columnstore. Because an entire segment has to be read for each column, we see a much greater number of reads and slower performance from a cold cache (0ms rowstore vs. 273ms columnstore). However, columnstore is down to 2ms with a warm cache; that's actually quite an impressive result given that there is no b-tree to seek into!

In the second test, we compute an aggregate for two columns across all rows. This is more along the lines of what columnstore is designed for, and we can see that columnstore has fewer reads (due to compression and not needing to access all columns) and dramatically faster performance (primarily due to batch mode execution). From a cold cache, columnstore executes in 4s vs 15s for rowstore. With a warm cache, the difference is a full order of magnitude at 282ms vs 2.8s.

SET STATISTICS TIME, IO ON
GO

-- Clear cache; don't do this in production!
-- I ran this statement between each set of trials to get a fresh read
--CHECKPOINT
--DBCC DROPCLEANBUFFERS
GO

-- Trial 1: CPU time = 0 ms,  elapsed time = 0 ms.
    -- logical reads 4, physical reads 4, read-ahead reads 0
-- Trial 2: CPU time = 0 ms,  elapsed time = 0 ms
    -- logical reads 4, physical reads 0, read-ahead reads 0
SELECT *
FROM dbo.test_row
WHERE id = 12345678
GO 2
-- Trial 1: CPU time = 15 ms,  elapsed time = 273 ms..
    -- lob logical reads 9101, lob physical reads 1, lob read-ahead reads 25756
-- Trial 2: CPU time = 0 ms,  elapsed time = 2 ms.  
    -- lob logical reads 9101, lob physical reads 0, lob read-ahead reads 0
SELECT *
FROM dbo.test_column
WHERE id = 12345678
GO 2

-- Trial 1: CPU time = 8441 ms,  elapsed time = 14985 ms.
    -- logical reads 264733, physical reads 3, read-ahead reads 263720
-- Trial 2: CPU time = 9733 ms,  elapsed time = 2776 ms.
    -- logical reads 264883, physical reads 0, read-ahead reads 0
SELECT AVG(id), SUM(col3)
FROM dbo.test_row
GO 2
-- Trial 1: CPU time = 1233 ms,  elapsed time = 3992 ms.
    -- lob logical reads 207778, lob physical reads 1, lob read-ahead reads 341196
-- Trial 2: CPU time = 1030 ms,  elapsed time = 282 ms. 
    -- lob logical reads 207778, lob physical reads 0, lob read-ahead reads 0
SELECT AVG(id), SUM(col3)
FROM dbo.test_column
GO 2

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.