Table Scan vs. Clustered Index Scan
-
Everyone agrees with presence of an index- or table-scan it means missing an index or the shaping of predicates is not really clear.
So far, anyone asked he replyed to me that nothing is worst like table scans. I think we must review this predicate.
Theoretically reading pages from the heap could be faster than reading cluster-index pages.
Background:
[font="Arial"] - Table scans or serial reads of a heap can be performed by scanning the IAM pages to find the extents that are holding pages for the heap. Because the IAM represents extents in the same order that they exist in the data files, this means that serial heap scans progress sequentially through each file.
- For a clustered index, the root_page column in sys.system_internals_allocation_units points to the top of the clustered index for a specific partition. SQL Server moves down the index to find the row corresponding to a clustered index key. To find a range of keys, SQL Server moves through the index to find the starting key value in the range and then scans through the data pages using the previous or next pointers. To find the first page in the chain of data pages, SQL Server follows the leftmost pointers from the root node of the index.
The storage engine scans the intermediate index page and builds a list of the leaf pages that must be read. The storage engine tschedules all the reads in key order.
- For sequential I/O performance, it is important to distinguish between allocation ordered and index ordered scans. An allocation ordered scan tries to read pages in the order in which they are physically stored on disk while an index ordered scan reads pages according to the order in which the data on those index pages is sorted. (Note that in many cases there are multiple levels of indirection such as RAID devices or SANS between the logical volumes that SQL Server sees and the physical disks. Thus, even an allocation ordered scan may in fact not be truly optimally ordered.) Although SQL Server tries to sort and read pages in allocation order even for an index ordered scan, an allocation ordered scan is generally going to be faster since pages are read in the order that they are written on disk with the minimal number of seeks. Heaps have no inherent order and, thus, are always scanned in allocation order. Indexes are scanned in allocation order only if the isolation level is read uncommitted (or the NOLOCK hint is used) and only if the query process does not request an ordered scan. Defragmenting indexes can help to ensure that index ordered scans perform on par with allocation ordered scans.[/font]
Measurement:
The test server is 2008 R2 developer edition, and not any load on it.
I chosed to test a table with 22278012 record, with following statistics of the Heap and Clustered index structure:
Index Type Rows Pages Size (MB) Frag. % Index Level
----------------------------------------------------------------------------------------
CLUSTERED 22278012 397090 3102 2,06 3
CLUSTERED 22278012 2382 18 6,85 2
CLUSTERED 22278012 23 1 95,65 1
CLUSTERED 22278012 1 0 0 0
HEAP 222780123877013028 0 0
The result-set of the query are written out into the tempdb which are placed in different physical disk.
First the buffer was cleared out with dbcc dropcleanbuffers command, to prevent reading pages from cache.
select * into #heap from [dbo].[heap_Table]
select * into #heap from [dbo].[Cl_Table]
Writing out the STATISTICS IO/TIME I got the below results:
Scan Type Logical r. Physical r. Read-ahead r. CPU time ms. Elapsed time ms.
----------------------------------------------------------------------------------------
Clustered 399475 878 399557 35084 57013
Clustered (nolock) 399496 59 399552 36286 50016
Heap 387701 28 387708 35319 49016
Heap (nolock) 387701 37 387708 35880 49141
Second time I let the pages to be kept in the buffer:
Scan Typecached_pages_countcached_pages_in_MB
----------------------------------------------------------------------------------------
Clustered2528421929
Clustered (nolock)2609361990
Heap3877092957
Heap (nolock)3877092957
Scan Type Logical r. Physical r. Read-ahead r. CPU time ms. Elapsed time ms.
----------------------------------------------------------------------------------------
Clustered 39947554 277101 38641 46511
Clustered (nolock) 399496188 397757 3686346564
Heap 3877010 17389 38376 44554
Heap (nolock) 38770124 31 3764343225
1. Clustered Index Scan - comitted reads
2. Clustered Index Scan - uncomitted reads (nolock)
Third I used the count methode to eliminate any writes.
select COUNT(ID) from [dbo].[heap_Table]
select COUNT(ID) from [dbo].[Cl_Table] with (index = IX_ID)
Scan Type Logical r. Physical r. Read-ahead r. CPU time ms. Elapsed time ms.
----------------------------------------------------------------------------------------
Clustered 3994750 0 5445 5438
Clustered (nolock) 3994960 0 54445464
Heap 3877010 0 5366 5386
Heap (nolock) 3877010 0 54135415
Empirical conclusion:
- In case of Clustered Index Scan we may read up the root + some pages from intermediate levels + leaf level
- The Clustered Index Scan with (nolock) hint we must read up all pages (root + all pages/all intermediate levels + leaf level)
- Scaning the Heap we have to read up the leaf level
From the first measurement table we can observe easily that the Table Scan is faster (with near 15%) than the Clustered Index Scan, however the CPU usage is higher with a bit but not significant (0,5%).
Reading up the top 1 million of rows I found that the Table San is still 11% faster than the Clustered Index Scan, moreover the Clustered Index Scan with (nolock) hint was 20% slower than Table Scan and the CPU usage increased with 12% too.
Very exciting the disk usage performance counter statistics in second methode.
In case of Heap all the pages are kept in the buffer and not any phisycal read observed in both of comitted or uncomitted read.
When we scan the Clustered Index we have lot of phisycal reads even if near all the pages are kept in memory.
In case of Clustered Index Scan in read uncomitted mode we read up again near all the pages from disk vainly we have the pages already in cache.
In read comitted mode is better the situation, because we read some from buffer, but still a lot from disk.
In third methode no significant difference between scan types (2%) which is expected since all the pages are scanned from buffer.
For me is very interesting why so different the operation principle of the scan methode depend we writing out the result or not.
-
kladna (3/2/2011)
From the above table we can observe easily that the Table Scan is faster (with near 15%) than the Clustered Index Scan, however the CPU usage is higher with a bit but not significant (0,5%).Reading up the top 1 million of rows I found that the Table San is still 11% faster than the Clustered Index Scan, moreover the Clustered Index Scan with (nolock) hint was 20% slower than Table Scan and the CPU usage increased with 12% too.
An interesting analysis, but you've taken the table as an independent entity and your results are assumed on always wanting the entire table loaded, neither of which is the common desire, especially amongst 22 million row tables.
Have you further pursued your testing to multiple tables being used (joins) and where clause mechanics? Even a HEAP with non-clustered indexing will still scan the leaf level in allocation method, so there are a number of options here.
- Craig FarrellNever stop learning, even if it hurts. Ego bruises are practically mandatory as you learn unless you've never risked enough to make a mistake.
For better assistance in answering your questions[/url] | Forum Netiquette
For index/tuning help, follow these directions.[/url] |Tally Tables[/url]
Twitter: @AnyWayDBA -
As you expected my article focusing strictly to this two entity, showing the background of the scaning method in different cases. I am surprized the difference between read comitted and uncomitted type of scans.
Also I meant to continue my measurment in multi-table query too, when is better to use Heap instead of Clustered Index, but that will be an another article.
-
these are the expected stats reason being .when we are calling every records (without any filer/where clause ) definitely Sql optimizer will go for table scan(or i would say clustered index scan ) because every records needs to be pulled out. but when we have any where clause condition or join operation definitely Clus index will give us better results.
-------Bhuvnesh----------
I work only to learn Sql Server...though my company pays me for getting their stuff done;-)
Viewing 4 posts - 1 through 3 (of 3 total)
You must be logged in to reply to this topic. Login to reply