I have found that when a table has both clustered and non-clustered indexes (on different columns), the leaf level non-clustered pages, instead of pointing to the data row, point to the node of the clustered index, from where another search is instituted to find the data row. What is the point of this extra level of indirection? If the clustered index has, say, 8 levels, then the indirection from the NCI leaf page to the CI root would have to traverse through these 8 levels to reach the data. Why not store the normal RID in the NCI leaf page so that we can access the data at once without going through the CI index structure?
Sql-server – Need for reaching data through clustered index with a non-clustered index
clustered-indexnonclustered-indexsql server
Related Solutions
To check your request I created 2 tables following this scheme:
- 7.9 million records representing balance information.
- an identity field counting from 1 to 7.9 million
- a number field grouping the records in about 500k groups.
The first table called heap got a non clustered index on the field group. The second table called clust got a clustered index on the sequential field called key and a nonclustered index on the field group
The tests were run on an I5 M540 processor with 2 hyperthreaded cores, 4Gb memory and 64-bit windows 7.
Microsoft SQL Server 2008 R2 (RTM) - 10.50.1600.1 (X64)
Apr 2 2010 15:48:46
Developer Edition (64-bit) on Windows NT 6.1 <X64> (Build 7601: Service Pack 1)
Update on 9 Mar 2011: I did a second more extensive benchmark by running the following .net code and logging Duration, CPU, Reads, Writes and RowCounts in Sql Server Profiler. (The CommandText used will be mentioned in the results.)
NOTE: CPU and Duration are expressed in milliseconds
- 1000 queries
- zero CPU queries are eliminated from the results
- 0 rows affected are eliminated from the results
int[] idList = new int[] { 6816588, 7086702, 6498815 ... }; // 1000 values here.
using (var conn = new SqlConnection(@"Data Source=myserver;Initial Catalog=mydb;Integrated Security=SSPI;"))
{
conn.Open();
using (var cmd = new SqlCommand())
{
cmd.Connection = conn;
cmd.CommandType = CommandType.Text;
cmd.CommandText = "select * from heap where common_key between @id and @id+1000";
cmd.Parameters.Add("@id", SqlDbType.Int);
cmd.Prepare();
foreach (int id in idList)
{
cmd.Parameters[0].Value = id;
using (var reader = cmd.ExecuteReader())
{
int count = 0;
while (reader.Read())
{
count++;
}
Console.WriteLine(String.Format("key: {0} => {1} rows", id, count));
}
}
}
}
End of Update on 9 Mar 2011.
SELECT performance
To check performanc numbers I performed the following queries once on the heap table and once on the clust table:
select * from heap/clust where group between 5678910 and 5679410
select * from heap/clust where group between 6234567 and 6234967
select * from heap/clust where group between 6455429 and 6455729
select * from heap/clust where group between 6655429 and 6655729
select * from heap/clust where group between 6955429 and 6955729
select * from heap/clust where group between 7195542 and 7155729
The results of this benchmark are for the heap:
rows reads CPU Elapsed
----- ----- ----- --------
1503 1510 31ms 309ms
401 405 15ms 283ms
2700 2709 0ms 472ms
0 3 0ms 30ms
2953 2962 32ms 257ms
0 0 0ms 0ms
Update on 9 Mar 2011:
cmd.CommandText = "select * from heap where group between @id and @id+1000";
- 721 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1001 69788 6368 -
Cpu 15 374 37 0.00754
Reads 1069 91459 7682 1.20155
Writes 0 0 0 0.00000
Duration 0.3716 282.4850 10.3672 0.00180
End of Update on 9 Mar 2011.
for the table clust the results are:
rows reads CPU Elapsed
----- ----- ----- --------
1503 4827 31ms 327ms
401 1241 0ms 242ms
2700 8372 0ms 410ms
0 3 0ms 0ms
2953 9060 47ms 213ms
0 0 0ms 0ms
Update on 9 Mar 2011:
cmd.CommandText = "select * from clust where group between @id and @id+1000";
- 721 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1001 69788 6056 -
Cpu 15 468 38 0.00782
Reads 3194 227018 20457 3.37618
Writes 0 0 0 0.0
Duration 0.3949 159.6223 11.5699 0.00214
End of Update on 9 Mar 2011.
SELECT WITH JOIN performance
cmd.CommandText = "select * from heap/clust h join keys k on h.group = k.group where h.group between @id and @id+1000";
The results of this benchmark are for the heap:
873 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1009 4170 1683 -
Cpu 15 47 18 0.01175
Reads 2145 5518 2867 1.79246
Writes 0 0 0 0.00000
Duration 0.8215 131.9583 1.9095 0.00123
The results of this benchmark are for the clust:
865 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1000 4143 1685 -
Cpu 15 47 18 0.01193
Reads 5320 18690 8237 4.97813
Writes 0 0 0 0.00000
Duration 0.9699 20.3217 1.7934 0.00109
UPDATE performance
The second batch of queries are update statements:
update heap/clust set amount = amount + 0 where group between 5678910 and 5679410
update heap/clust set amount = amount + 0 where group between 6234567 and 6234967
update heap/clust set amount = amount + 0 where group between 6455429 and 6455729
update heap/clust set amount = amount + 0 where group between 6655429 and 6655729
update heap/clust set amount = amount + 0 where group between 6955429 and 6955729
update heap/clust set amount = amount + 0 where group between 7195542 and 7155729
the results of this benchmark for the heap:
rows reads CPU Elapsed
----- ----- ----- --------
1503 3013 31ms 175ms
401 806 0ms 22ms
2700 5409 47ms 100ms
0 3 0ms 0ms
2953 5915 31ms 88ms
0 0 0ms 0ms
Update on 9 Mar 2011:
cmd.CommandText = "update heap set amount = amount + @id where group between @id and @id+1000";
- 811 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1001 69788 5598 811
Cpu 15 873 56 0.01199
Reads 2080 167593 11809 2.11217
Writes 0 1687 121 0.02170
Duration 0.6705 514.5347 17.2041 0.00344
End of Update on 9 Mar 2011.
the results of this benchmark for the clust:
rows reads CPU Elapsed
----- ----- ----- --------
1503 9126 16ms 35ms
401 2444 0ms 4ms
2700 16385 31ms 54ms
0 3 0ms 0ms
2953 17919 31ms 35ms
0 0 0ms 0ms
Update on 9 Mar 2011:
cmd.CommandText = "update clust set amount = amount + @id where group between @id and @id+1000";
- 853 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1001 69788 5420 -
Cpu 15 594 50 0.01073
Reads 6226 432237 33597 6.20450
Writes 0 1730 110 0.01971
Duration 0.9134 193.7685 8.2919 0.00155
End of Update on 9 Mar 2011.
DELETE benchmarks
the third batch of queries I ran are delete statements
delete heap/clust where group between 5678910 and 5679410
delete heap/clust where group between 6234567 and 6234967
delete heap/clust where group between 6455429 and 6455729
delete heap/clust where group between 6655429 and 6655729
delete heap/clust where group between 6955429 and 6955729
delete heap/clust where group between 7195542 and 7155729
The result of this benchmark for the heap:
rows reads CPU Elapsed
----- ----- ----- --------
1503 10630 62ms 179ms
401 2838 0ms 26ms
2700 19077 47ms 87ms
0 4 0ms 0ms
2953 20865 62ms 196ms
0 4 0ms 9ms
Update on 9 Mar 2011:
cmd.CommandText = "delete heap where group between @id and @id+1000";
- 724 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 192 69788 4781 -
Cpu 15 499 45 0.01247
Reads 841 307958 20987 4.37880
Writes 2 1819 127 0.02648
Duration 0.3775 1534.3383 17.2412 0.00349
End of Update on 9 Mar 2011.
the result of this benchmark for the clust:
rows reads CPU Elapsed
----- ----- ----- --------
1503 9228 16ms 55ms
401 3681 0ms 50ms
2700 24644 46ms 79ms
0 3 0ms 0ms
2953 26955 47ms 92ms
0 3 0ms 0ms
Update on 9 Mar 2011:
cmd.CommandText = "delete clust where group between @id and @id+1000";
- 751 Rows have > 0 CPU and affect more than 0 rows
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 144 69788 4648 -
Cpu 15 764 56 0.01538
Reads 989 458467 30207 6.48490
Writes 2 1830 127 0.02694
Duration 0.2938 2512.1968 24.3714 0.00555
End of Update on 9 Mar 2011.
INSERT benchmarks
The last part of the benchmark is the execution of insert statements.
insert into heap/clust (...) values (...), (...), (...), (...), (...), (...)
The result of this benchmark for the heap:
rows reads CPU Elapsed
----- ----- ----- --------
6 38 0ms 31ms
Update on 9 Mar 2011:
string str = @"insert into heap (group, currency, year, period, domain_id, mtdAmount, mtdAmount, ytdAmount, amount, ytd_restated, restated, auditDate, auditUser)
values";
for (int x = 0; x < 999; x++)
{
str += string.Format(@"(@id + {0}, 'EUR', 2012, 2, 0, 100, 100, 1000 + @id,1000, 1000,1000, current_timestamp, 'test'), ", x);
}
str += string.Format(@"(@id, 'CAD', 2012, 2, 0, 100, 100, 1000 + @id,1000, 1000,1000, current_timestamp, 'test') ", 1000);
cmd.CommandText = str;
- 912 statements have > 0 CPU
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1000 1000 1000 -
Cpu 15 2138 25 0.02500
Reads 5212 7069 6328 6.32837
Writes 16 34 22 0.02222
Duration 1.6336 293.2132 4.4009 0.00440
End of Update on 9 Mar 2011.
The result of this benchmark for the clust:
rows reads CPU Elapsed
----- ----- ----- --------
6 50 0ms 18ms
Update on 9 Mar 2011:
string str = @"insert into clust (group, currency, year, period, domain_id, mtdAmount, mtdAmount, ytdAmount, amount, ytd_restated, restated, auditDate, auditUser)
values";
for (int x = 0; x < 999; x++)
{
str += string.Format(@"(@id + {0}, 'EUR', 2012, 2, 0, 100, 100, 1000 + @id,1000, 1000,1000, current_timestamp, 'test'), ", x);
}
str += string.Format(@"(@id, 'CAD', 2012, 2, 0, 100, 100, 1000 + @id,1000, 1000,1000, current_timestamp, 'test') ", 1000);
cmd.CommandText = str;
- 946 statements have > 0 CPU
Counter Minimum Maximum Average Weighted
--------- ------- ---------- ------- ---------
RowCounts 1000 1000 1000 -
Cpu 15 2403 21 0.02157
Reads 6810 8997 8412 8.41223
Writes 16 25 19 0.01942
Duration 1.5375 268.2571 6.1463 0.00614
End of Update on 9 Mar 2011.
Conclusions
Although there are more logical reads going on when accessing the table with the clustered & the nonclustered index (while using the nonclustered index) the performance results are:
- SELECT statements are comparable
- UPDATE statements are faster with a clustered index in place
- DELETE statements are faster with a clustered index in place
- INSERT statements are faster with a clustered index in place
Of course my benchmark was very limited on a specific kind of table and with a very limited set of queries, but I think that based on this information we can already start saying that it is virtually always better to create a clustered index on your table.
Update on 9 Mar 2011:
As we can see from the added results, the conclusions on the limited tests were not correct in every case.
The results now indicate that the only statements which benefit from the clustered index are the update statements. The other statements are about 30% slower on the table with clustered index.
Some additional charts where I plotted the weighted duration per query for heap vs clust.
As you can see the performance profile for the insert statements is quite interesting. The spikes are caused by a few data points which take a lot longer to complete.
End of Update on 9 Mar 2011.
I think the problem here is a difference in terminology.
The "number of writes" that's usually referred to is the number of object accesses, rather than the number of pages that get touched by the physical operation.
The reason why that's usually used as a metric in discussion is because it's a more "stable" and meaningful number to talk about. As we're getting into here, the number of pages touched by an INSERT statement for even a single row depends on many factors, so it's not a very useful quantity outside your own environment and situation.
The one thing I would pick at from the article quote is this (emphasis mine):
One write for inserting the row, and one write for updating the non-clustered index.
This may be confusing. Inserting a row into the base table would involve an insert to the base table, and also an insert into each nonclustered index (ignoring special index features), not an update.
So if a record has to be updated, say the value 1 has to be updated to 7, won't the update need to be applied to both the key in the clustered index top node (this may, in cases, cause a re-structuring of the entire structure) and the corresponding value in the record in the leaf-page?
Yes, assuming the column that was updated is in the index key. However, this is still a single object access, and hence a "single write."
Related Question
- Sql-server – Finding exact row in Clustered index leaf page
- Sql-server – Locating row in index page
- Sql-server – Clustered index additional storage
- Sql-server – What happens when you delete a Clustered index and Non Clustered Index once created
- Sql-server – Non clustered index leaf level
- Sql-server – Is the default clustered index file physically the same file as the file containing the data records
Best Answer
The reason for this is that the "fixed" physical location of your row - the RID (or row identifier) might (and will!) change over time - think page splits that occur when a row needs to be inserted into a table on a page that's already full.
Updating those RIDs in all the nonclustered indices that exist on a given table is quickly becoming both a hassle, and a huge performance killer. You might have 5, 10, 20 nonclustered indices on your table, and SQL Server would have to scan all those indices (basically scanning the whole index, all rows in the index, and that 10, 20 times) and update all RIDs.... that's just not practical - very quickly so.
If you store the value of the clustering index as a "row pointer" instead, that value should typically never change - and most definitely it doesn't need to be updated every time a page is split. Yes, it does involve a second index seek operation - the key lookup - but for simple scenarios, retrieving a single row or a few rows, that's still much more efficient than anything else.