Tips, Tricks, and Advice from the SQL Server Query Processing Team

Mystery of memory fraction in Showplan XML

2010-03-12T02:34:00Z

If you examine Showplan XML, you may find an entry like <MemoryFractions Input="1" Output="1" /> on some operators . They are internal parameters used by query memory grant to save overall memory usage, and “fraction” means how much this operator is allowed to use from “total additional memory” (please see previous post for definition on additional memory).

For example, let’s consider the following hypothetical plan:

SORT – HJ (hash join) – complex build side child
|--- probe side child

Since server knows HJ does not need to access its build-side once hash table is built, it can delay creation of sort table until HJ is ready to output joined rows. Then SORT can reuse memory previously used by HJ build side child. This is done by observing input (building hash table) to output (probing) phase transition by HJ.

So what’s the practical use of these values? Not much beyond confirming memory grant usage, if the fraction is comparable to 1. If it is very small (say 0.1), however, the query may be susceptible for bad cardinality estimate. If you experience bad performance due to disk spilling and see such small number in Showplan XML, then it may be a good starting point for investigation.

--- Jay Choe, SQL Server Engine

Understanding SQL server memory grant

2010-02-16T18:53:00Z

This article describes how query memory grant works in Microsoft SQL Server. It applies to both SQL2005 and SQL2008. Written by Jay Choe, Microsoft SQL Server Engine.

--------------------------------------------------------------------------------

Query memory grant (a.k.a. query work buffer) is a part of server memory used to store temporary row data while sorting and joining rows. It is called "grant" because the server requires those queries to "reserve" before actually using memory. This reservation improves query reliability under server load, because a query with reserved memory is less likely to hit out-of-memory while running, and the server prevents one query from dominating entire server memory. In the following sections, I will describe how SQL server estimates query memory requirement, and how it throttles memory grants when multiple queries are competing.

When SQL server receives a user query, it follows well-defined steps to produce a result for user. First, it creates a compiled plan, which is a set of logical instructions such as how to join rows. Next, it creates an execution plan based on the compiled plan. This execution plan contains instructions with all logical references translated to actual objects, and tracking mechanisms for query execution. Finally, the server starts execution from the top of instruction tree. Creating a compiled plan is expensive because the server needs to find the optimal plan out of hundreds of potential candidates. Distinguishing between compile and execution helps overall server performance because each compiled plan can be cached and shared among multiple execution plans. The memory grant estimate follows the same overall sequence. It has parameters saved in compiled plan, and a mechanism to calculate actual grant size at execution time.

Memory consumers
Now that I briefly covered the lifetime of query execution, I would like to show where query memory grant fits in overall server memory usage. As hinted in the previous paragraph, a successful query execution involves 3 major memory consumers: compile, cache, and memory grant.

Compile (query optimization): Building and searching the most optimal plan out of hundreds candidates typically requires significant amount of memory. The lifetime of this usage is typically short because optimizer releases memory as soon as optimal plan is found. Lack of available memory would cause delays in compile, and potentially inefficient (slow) plans.
Cache: Finding optimal plan is costly in terms of CPU and memory usage. SQL server tries to store compiled plans in caches for later reuse. Lifetime of this memory usage is long-term. Lack of cache memory would cause more unnecessary re-compiles.
Memory grant: This memory is used to store temporary rows for sort and hash join. The lifetime of memory grant is the same as the lifetime of query. Lack of available memory grant causes a query to use hard disk, which affects query performance.

SQL Server maintains the balance between these 3 consumers with internal facility called "memory broker". Based on the usage and physical memory available, the memory broker sets the usage limit and tells each component to trim its memory usage if it anticipates a shortage. Generally, a well behaving server would have about the same contributions from these consumers.

Grant parameters
When SQL server creates a compiled plan, it calculates two memory grant parameters called "required memory" and "additional memory".

Required memory: Minimum memory needed to run sort and hash join. It is called required because a query would not start without this memory available. SQL server uses this memory to create internal data structures to handle sort and hash join.
Additional memory: Amount needed to store all temporary rows in memory. This depends on the cardinality estimate (expected number rows and average size of row). This is called additional because a query can survive lack of such memory by storing part of temporary rows on hard disk. A query is not guaranteed to have the full amount if the total exceeds preset limit.

For example, let's consider a simple query which needs to sort 1 million rows of 10 bytes each in size. The required memory for this query is 512KB because this is the minimum amount SQL server needs to construct internal data structures to handle one sort. Since it would take 10MB to store all rows, the additional memory would be 10MB (slightly higher when overhead is included). This calculation becomes complicated if the compiled plan has multiple sorts and joins because SQL server also considers the lifetime of each operator for more efficient memory usage. You would generally see smaller estimate than the sum of all sorts and joins. If you are interested in the relative memory usages among operators, you can check out <MemoryFractions> tag in Showplan XML. The following sections show how these parameters are used when calculating grant size at runtime.

DOP dependence
If SQL server has more than 1 CPU, it can run a query in parallel mode for improved performance by sharing work among parallel workers. These workers run independent of each other, and use "parallelism operator (a.k.a. exchange)" to transfer processed rows. This parallel mode increases memory usage because each worker needs its own copy of sort or hash join, and the parallelism operator needs buffers for temporary storage of transferred rows. Since DOP N would use N parallel workers, the query would need N times more required memory. On the other hand, the total number of rows to handle (and memory cost to store them) does not change with DOP. This means that the additional memory would stay the same regardless of DOP setting. Starting with SQL 2008, the buffer memory used by parallelism operator is also counted as a required part of memory grant.

Memory grant process
In the previous section, we discussed how parallelism affects the query memory requirement. In this section, we will discuss how SQL server takes server memory and number of concurrent queries into considerations. The server needs to consider such dynamic factors to avoid committing memory beyond its physical limit. This is done in 2 distinct steps. First, the server calculates how much memory to grant for given query. Then it uses the internal facility called Resource Semaphore to reserve actual memory, or throttle if too many queries ask for memory. First, the following steps show how the request size is determined.

The server decides parallelism (DOP) based on the plan and the server state.
The server checks if memory grant is needed not. If not needed, the server can start the query immediately. For example, a simple serial query without "ORDER BY" or "GROUP BY" may not need memory grant.
The server calculates the memory limit for one query. By default, this is 25% (20% on 32bit SQL 2005) of total query memory (which is set by memory broker as about 90% of server memory). This per-query limit helps to prevent one query from dominating the whole server. This percentage is configurable on SQL 2008.
The server calculates the ideal query memory as required*DOP + additional (+ exchange on SQL 2008). This is amount a query would like to have, based on its cardinality estimate.
The server checks if the ideal memory exceeds the per-query limit. If it does, then the server reduces the additional memory until the total fits within the limit. This revised size is called requested memory.The server asks Resource Semaphore to grant the requested memory.

Resource Semaphore
Resource Semaphore is responsible for satisfying memory grant requests while keeping overall memory grant usages within the server limit.

Resource Semaphore allows a query to reserve memory only if there is enough free memory available. Otherwise, a requesting query is forced to wait in a queue.
When Resource Semaphore receives a new request, it first checks if any query is waiting. If it finds one, it puts the new query into a queue for fairness because the wait queue is designed as first-come-first-served basis.
Resource Semaphore checks for waiting query in its queue. If it finds one, it puts the new query into a wait queue for fairness, because the wait queue is designed as first-come-first-served basis with small weight to favor small queries.
Resource Semaphore makes a grant attempt if it does not find waiting query, or when existing query returns memory.
Resource Semaphore attempts to grant memory when there is no waiting query, or when a query returns memory.
Grant attempt is made when there is no waiting query, or when an existing query returns reserved memory.
If it finds a waiting query, it puts the current query into a waiting queue for fairness.
If it does not find any waiting query, it then checks available free memory.
If it finds enough free memory, then the requested memory is granted and the query can start running.
If it does not find enough free memory, then it puts the current query into the waiting queue.
Resource Semaphore wakes up queries in the waiting queue when enough free memory becomes available.

Debugging memory grant related issues
SQL server supplies a few dynamic management views (DMV) to help investigate memory grant related issues. Please refer BOL (books-on-line) for detailed information on DMVs discussed below.

sys.dm_exec_query_resource_semaphores
This DMV shows the current status of Resource Semaphore mentioned in the previous section. On SQL 2005, you will find 2 of them. One with non-null max_target_memory_kb column is called Regular Resource Semaphore, and the other is called Small Resource Semaphore. As its name implies, Regular Resource Semaphore is used by all queries under normal condition, while Small Resource Semaphore is used by small size queries (less than 5 MB) when they have to wait (see step 6 in the previous section). This would improve response time of small queries which are expected to finish very fast. Note max_target_memory_kb column shows the server memory limit used in step 2 of previous section.

sys.dm_exec_query_memory_grants
This DMV shows all queries consuming memory grants including those waiting in Resource Semaphore queue. Waiting queries would have null grant_time column. Resource Semaphore uses internal query cost estimate to prioritize memory grants, and is_next_candidate column shows which query to wake up when memory is available.

sys.dm_os_wait_stats
This DMV shows wait statistics of all server objects. Memory grant uses "RESOURCE_SEMAHORE" wait type. If you see significant waits on this wait type, you may have an issue with big queries.

Sample queries
The following sample queries show how memory grant DMV are used.

Find all queries waiting in the memory queue:

SELECT * FROM sys.dm_exec_query_memory_grants where grant_time is null

Find who uses the most query memory grant:

SELECT mg.granted_memory_kb, mg.session_id, t.text, qp.query_plan
FROM sys.dm_exec_query_memory_grants AS mg
CROSS APPLY sys.dm_exec_sql_text(mg.sql_handle) AS t
CROSS APPLY sys.dm_exec_query_plan(mg.plan_handle) AS qp
ORDER BY 1 DESC OPTION (MAXDOP 1)

Search cache for queries with memory grants:

SELECT t.text, cp.objtype,qp.query_plan
FROM sys.dm_exec_cached_plans AS cp
JOIN sys.dm_exec_query_stats AS qs ON cp.plan_handle = qs.plan_handle
CROSS APPLY sys.dm_exec_query_plan(cp.plan_handle) AS qp
CROSS APPLY sys.dm_exec_sql_text(qs.sql_handle) AS t
WHERE qp.query_plan.exist('declare namespace n="http://schemas.microsoft.com/sqlserver/2004/07/showplan"; //n:MemoryFractions') = 1

Understanding SQL Server Fast_Forward Server Cursors

2009-08-13T03:38:00Z

SQL Server's server cursor model is a critical tool to many application writers. Fast_forward cursors are very popular as an alternative to read_only forward_only cursors, but their inner workings are not well-publicized. So I thought I'd give it a go.

Background

A server cursor is a cursor managed by SQL Engine. It consists of a query execution and some runtime state, including a current position. SQL clients can use a server cursor to fetch results of a query one-at-a-time, or in batches of N-at-a-time, rather than by the usual all-at-a-time, firehose, default query result set. Many applications and client libraries have iteration models, but server cursors are the only way for SQL Server engine to do incremental query processing.

Server cursors' popularity is derived from three advantages:

1. Matching very well the iterative model of programming that non-database programmers are most comfortable with. This advantage is more perceived than actual, and does more harm than good. Books Online puts it well:

If the data that is to be retrieved will be consumed all at once, and there is no need for positioned updates or scrolling, default result sets are recommended.

2. Reducing unnecessary processing of unused data when a client application stops reading data early. Because in many cases (though not all), server cursor processing is incremental and on-demand, server load and data transmission can be reduced.

3. Mapping very well to screen-at-a-time or window-at-a-time viewing and scrolling through large data sets. There is simply no better way to tell SQL Server to "give me the next row(s)."

SQL Server engine has four server cursor models: static, keyset, dynamic, and fast_forward. [Apologies for not using the ADO.NET terminology -- it tends to confuse me. How various settings on ADO.NET and ODBC commands translate to server cursor models is a topic for another time.] There is lots of further background on the models and their differences in SQL Server 2008 Books Online's section on cursors. But my concern here is to demystify fast_forward cursors, about which Books Online say very little, perhaps worse than nothing:

FAST_FORWARD

Specifies a FORWARD_ONLY, READ_ONLY cursor with performance optimizations enabled.

What's a fast_forward cursor in a nutshell?

It's a cursor model equivalent to read_only, forward_only that compiles to a static-like or dynamic-like cursor plan. It does not downgrade to other cursor models, like dynamic and keyset do.

Aren't fast_forward cursors redundant? We already have read_only forward_only cursors. Why do we need them?

Fast_forward is redundant. Read_only forward_only cursors are sufficient for many apps. But some apps got poor query plans. Read_only forward_only cursors are dynamic cursors, and dynamic cursors use a dynamic plan if one is available. The reason this is a problem is that sometimes, the best dynamic plan is much, much worse than a static one. Fast_forward cursors take a more balanced approach, choosing a static plan if it looks better.

When should I use fast_forward vs. read_only forward_only?

On balance, fast_forward is a little better. [1][1] However, performance testing should be done before a final decision is made for a particular application. This is because the way the decision is made to use a dynamic-like or a static-like plan is fundamentally different (see below). Either way, index tuning and plan hints may be an important part of tuning your cursors, whichever cursor model you choose.

What's a dynamic plan?

A dynamic plan can be processed incrementally. In SQL Server we do this by serializing the state of the query execution into what we call a marker. Later, we can build a new query execution tree, use the marker to reposition each operator. Moreover, a dynamic plan can move forwards and backwards relative to its current position. Dynamic plans are used by both dynamic and some fast_forward cursors.

A dynamic plan consists only of dynamic operators -- operators that support markers and moving forwards and backwards. This corresponds closely, but not exactly, to the query processing notion of streaming operators (vs. stop-and-go). But not every streaming operator is dynamic. In SQL Server, dynamic means:

1. The operator can be repositioned to its current position using a marker, or to a relative position (either next or previous) from its current one.

2. The operator's state has to be small, so the marker can be small. No row data can be stored in the operator. In particular, no sort table, hash table, or work table. Not even one row can be stored, since a single row can be very large.

Without a dynamic plan, the cursor would need temporary storage to keep the query result set (or keyset thereof). A dynamic plan does no such thing! However, certain operators are disqualified -- hash join, hash agg, compute sequence, and sort, for example. This leads to sub-optimal plans.

When is a dynamic plan worse than a static one?

In some queries, like those that use row_number, a dynamic plan is simply not available. The trouble comes when there are both dynamic and static alternatives. This can happen when an operator, e.g. join, has dynamic (nested loops) and non-dynamic (hash) implementations. This can also happen when some indexes support interesting orders and some do not.

Here is an example of the latter drawn from a real customer case. Table ORDERS has an index on DATE and an index on SUBTOTAL (clustered/non-clustered is not relevant for the example). The application wants to see very large orders:

SELECT DATE, SUBTOTAL, ORDERID, CUSTOMERID

FROM ORDERS where SUBTOTAL > 10000

ORDER BY DATE

For argument's sake, say there are 100 orders this large out of a table of 100 million orders. A dynamic cursor cannot sort, so it must use the DATE index, touching all the rows, filtering on SUBTOTAL. A static cursor has the option of seeking the SUBTOTAL index for the range >10000, sorting the rows, and storing them in the cursor.

How do fast_forward cursors solve this problem?

Fast_forward cursors under certain conditions[2][2] choose the cheaper of the best static and best dynamic plans. In an extreme case like the above, it can use a static plan. It should be the best of both worlds.

What's the hitch?

The optimizer relies on cost estimates to decide on the cursor model. Sometimes, when it picks static rather than dynamic, it will simply be wrong, and the application will be very slow. There are three reasons why this happens.

1. The cost of a query depends on the "row goal": the number of rows estimated to be fetched. In many cursor applications, user are viewing a screen with a few rows and a next button or scroll bar. If the user is likely to scroll through all the data, the plan cost is different than if the user is likely to quit after viewing the first page. We have to pick a row goal at compile time to estimate a cost, but the user's behavior is unknown at that time. Typically, static plans are not sensitive to row goal, while dynamic plans are critically sensitive. In the case above, the dynamic plan cost scales linearly with the row goal (and when there are joins, it's super-linear). That makes the dynamic plan superior by an arbitrary factor for low row goals, and inferior by an arbitrary factor for high row goals. The factors depend on query and data parameters (such as the 10000, and the data distribution, in the example). Application writers can control row goals by specifying OPTION (FAST <N>) on their cursor queries, but often they cannot predict the correct value either.

2. Parameterized queries are compiled based on the parameters at hand, and then reused for different parameters. In the example above, it makes a big difference whether the cut-off is at 10000 or at 10. This is no different from non-cursor queries.

3. The choice of plan is subject to estimation errors, which can be extreme, especially when combined with the previous factors. A typical flavor of disaster involves two indexes like in the example, where the server estimates that the predicate is super selective, and therefore will scan half the table before finding the first qualifying row, therefore picking the static plan. However, if there is a qualifying row on the first page, and the user only wants the first row, then the dynamic plan is very, very cheap.

What can an application developer do?

For canned queries, via trial and error, you can use hints to get you the cursor plan you want. For query generators, it is much more difficult, and there's no one solution. You do have a few levers at your disposal, though.

1. Picking a neutral value for OPTION (FAST <N>).

2. Avoiding plan reuse for different parameters by calling sp_cursorprepare or by using OPTION (RECOMPILE).

3. Avoiding ORDER BY in cursors when not really needed.

4. Use equality predicates and multi-column indexes to support interesting orders efficiently. Modifying the example slightly, if we had a computed column SIZE with values {S,M,L,XL} for ranges of SUBTOTAL, we could query WHERE SIZE = 'XL' ORDER BY DATE, and index on <SIZE, DATE>. In general, extend the index columns to include the equality columns followed by the ordering columns.

By Marc Friedman, 8/2009

Marc.friedman@donotreply.microsoft.com

Note to self: ref: VSTS# 273442

[1][1] For SQL Server 2005 users, it's worth mentioning that there is a performance enhancement to fast-forward prefetched I/O in SP3 Cumulative Update 5 that brings fast_forward up to parity. SQL Server 2008 and beyond, this is a non-issue.

[2][2] In situations where the dynamic plan looks promising, the cost comparison may be heuristically skipped. This occurs mainly for extremely cheap queries, though the details are esoteric. In that case no difference between dynamic and fast_forward will be noticed.

Distinct Aggregation Considered Harmful

2008-09-22T23:47:00Z

Distinct aggregation (e.g. select count(distinct key) …) is a SQL language feature that results in some very slow queries. It’s particularly frustrating that you can take a perfectly efficient query with multiple aggregates, and make that query take forever just by adding a distinct keyword to one of the aggregates. For instance, it often makes logical sense to compute distinct counts “along the way” while you are computing other aggregates. Recently, I've seen a number of customer queries like this:

select

sum(salary),

max(salary),

count(employeeid),

count(distinct employeeid),

count(distinct salarygrade)

In practice, however, SQL Server 2008 does not compute these distinct aggregates “along the way.” Mixing one or more distinct aggregates with non-distinct aggregates in the same select list, or mixing two or more distinct aggregates, leads to spooling and rereading of intermediate results – which can be more expensive than computing the distinct aggregates in a separate query! In the worst case, such as the fragment of code above, this is sort of inevitable. The SQL Server 2008 query processor cannot compute aggregates on a stream without destroying the stream. So computing the distinct aggregate consumes the stream, which has to be recomputed for the other aggregates.

Fortunately, for select lists with only one distinct aggregate, you can rewrite the input SQL in a way that does not have this problem. SQL Server 2008’s optimizer does not do this rewrite for you. (Disclaimers: all the aggregates have to be algebraic [Gray et al 1996], and no guarantees are made that the behavior of the rewrite is exactly the same, particularly when there are arithmetic overflows.)

Getting Rid of Distinct Aggregates

The code below shows an example rewrite using the AdventureWorksDW database distributed with SQL Server 2008. The rewrite gives me 5x speedup on my desktop even on this small database! What’s going on is that we are breaking the aggregation into two aggregation steps. First, we build an intermediate result called PartialSums. We group together all the values for the distinct aggregate (adding the distinct aggregate’s column to the GROUP BY list), while building partial aggregations for all the non-distinct aggregates (in this example, just count(*)). Then in the second step, we use the original GROUP BY list, which in this example is empty, and collect the partial aggregates together. Note how the aggregate functions used depend on the initial aggregates: count becomes count, followed by sum. Note that in sum(1), the 1 is actually a constant number value, not a column reference.

As far as I can tell, this rewrite is never worse than the spooled plan. It parallelizes better, uses tempdb better, and does less logical I/O.

set statistics profile on

set statistics time on

use adventureworksdw

-- Use a distinct aggregate and a normal aggregate in the same select list

-- over some complex (one or more joins) derived table

with FISinFRS as (

select * from factinternetsales FIS

where ProductKey in

(select ProductKey from FactResellerSales)

)

select

count(*) as CountStar,

count(distinct ProductKey) as CountProductKeys

from FISinFRS

-- Now use partial aggregations in a new derived table

-- This is equivalent, but SQL Server 2008 does not do this transformation

-- for you

with FISinFRS as (

select * from factinternetsales FIS

where ProductKey in

(select ProductKey from FactResellerSales)

)

, PartialSums as (

select

count(*) as CountStarPartialCount

from FISinFRS

group by ProductKey

)

select

sum(CountStarPartialCount) as CountStar,

sum(1) as CountProductKeys

from PartialSums

References

[Gray et al 1996] Data cube: A relational aggregation operator generalizing group-by, cross-tab, and sub-totals (1996), ftp://ftp.research.microsoft.com/pub/tr/tr-95-22.pdf.

By Marc Friedman, 9/2008

Marc.friedman@donotreply.microsoft.com

Store Statistics XML in database tables using SQL Traces for further analysis.

2007-06-01T19:51:00Z

Since SQL Server 2005, query plan as well as statistics of query execution can be captured in XML format. Also, SQL Server 2005 has XQuery support to directly query XML document. By combining these two new features, users can analyze the query plans using queries.

However, in SQL Server, there is no easy way to capture the statitics XML into a table. Fortunately, there are SQL traces provided by SQL Server to capture the showplan XML and statistics XML information into trace files and loaded into tables.

Note: The only limitation is the 128 level of nesting levels supported by XML data type in SQL 2005. In that case, you have to write client code to parse the query plan, which going to be a very complex query plan.

Here is a small example using SQL traces to store the statistics XML and extract the estimated rows and actual rows.

/*Using Traces to Capture Statistics XML*/

declare @trace_id int

declare @trace_file nvarchar(200)

select @trace_file = 'c:\temp\test_stats_' + cast(newid() as varchar(100))

-- using trace table.

exec sp_trace_create @trace_id output,

@tracefile=@trace_file

-- capture statistics-xml, textdata, on

exec sp_trace_setevent @trace_id, 146, 1, 1

-- start

exec sp_trace_setstatus @trace_id, 1

-- test statement.

select * from sys.objects

-- stop

exec sp_trace_setstatus @trace_id, 0

-- close

exec sp_trace_setstatus @trace_id, 2

-- load trace files into table

if object_id('temp_trc') is not null

drop table temp_trc

select *

into temp_trc

from fn_trace_gettable(@trace_file + '.trc', default)

-- look at the captured stats xml

declare @plan xml

select @plan=cast(textdata as xml)

from temp_trc

where eventclass = 146;

-- collect the actual and also estimate stats.

with XMLNAMESPACES ('http://schemas.microsoft.com/sqlserver/2004/07/showplan' as sql)

select

ro.relop.value('@NodeId', 'int') NodeId,

ro.relop.value('@PhysicalOp', 'nvarchar(200)') PhysicalOp,

ro.relop.value('@LogicalOp', 'nvarchar(200)') LogicalOp,

(ro.relop.value('@EstimateRows', 'float')

* (ro.relop.value('@EstimateRewinds', 'float')

+ ro.relop.value('@EstimateRebinds', 'float')

+ 1.0)) EstimateRows,

case

when root_actual.ActualRows = 0

then null

else root_actual.ActualRows

end ActualRows,

cast(ro.relop.exist('*/sql:RelOp') as bit) IsNotLeaf

from @plan.nodes('//sql:RelOp') as ro(relop)

cross apply (

select sum(rti.info.value('@ActualRows', 'float')) ActualRows

from ro.relop.nodes('sql:RunTimeInformation/sql:RunTimeCountersPerThread') as rti(info)

) root_actual;

The output of the estimate rows and actual rows is given below:

NodeId	PhysicalOp	LogicalOp	EstimateRows	ActualRows	IsNotLeaf
0	Nested Loops	Left Outer Join	52	52	1
1	Nested Loops	Left Outer Join	52	52	1
2	Filter	Filter	52	52	1
3	Compute Scalar	Compute Scalar	52	NULL	1
4	Clustered Index Scan	Clustered Index Scan	52	52	0
13	Clustered Index Seek	Clustered Index Seek	52.000031	NULL	0
14	Clustered Index Seek	Clustered Index Seek	52	52	0

Index Build strategy in SQL Server - Part 4-2: Offline Serial/Parallel Partitioning (Non-aligned partitioned index build)

2007-05-14T07:15:00Z

Source Partitioned
While the table is partitioned, we may want to change the way it is partitioned when building the new index. For example, by using the same partition function and scheme, the new index can be partitioned on different columns than the original table.

Create table t (c1 int, c2 int) on ps(c2)

…….

Create clustered Index idx_t on t(c1) on ps(c1)

The serial plan looks like follows.
Index Insert
   |
Sort
   |
NL (Nested Loop)
/    \
CTS Scan
CTS is the Constant Table Scan. It scans each partition one by one and provides partition ID to the inner side (the lower part in graphics showplan) of NL. The inner side of NL scans the corresponding partition and sends data to the sort iterator. From there, it is exactly the same as source non-partitioned scenario. Not surprisingly, the memory and disk space requirement is the same too.

In the case of parallel plan, we have
      X (Distribute Streams)
       |
     Index Insert
       |
     Sort
       |
       X (Repartition Streams)
       |
      NL
     /   \
   X    Scan
/
CTS
The operator above CTS is a Gather Streams operator, meaning it has one producer and many consumers. There is no parallelism below this operator. Between the Gather Streams and the Repartition Streams, each worker is assigned (Number of Source's Partitions)/(Degree of Parallelism) number of source partitions. The source is only scanned once in total.

The Repartition Streams operator splits the query plan into two parallelism sections. Between the top-level Distribute Streams and the Repartition Streams, we have a different set of workers than the worker set below Repartition Streams. Each worker is assigned (Number of Target Index Partitions)/(Degree of Parallelism) target partitions. The target index partition information is pushed down to the Repartition Streams which redirects data to different sort tables based on target partition location. The rest, i.e. how the sort and index building works, is the same as the parallel plan in the source non-partitioned case. Again, the memory and disk space requirement is the same as in the source non-partitioned case.

Index Build strategy in SQL Server - Part 4-1: Offline Serial/Parallel Partitioning (Non-aligned partitioned index build)

2007-05-08T21:56:00Z

Recall that in the previous posts on index build, we defined "aligned" as the case when base object and in-build index use the same partition schema, and "non-aligned" to be the case when heap and index use different partition schemes, or the case when heap is not partitioned. In this post, we will talk about the two scenarios of non-aligned partitioned index build, source partitioned and source not partitioned.

Source Not Partitioned
Consider the following query.

Create Partition Function pf (int)

as range right for values (1, 100, 1000)

Create Partition Scheme ps as Partition pf

ALL TO ([PRIMARY])

Create table t (c1 int, c2 int) –the table is created on the primary filegroup by default

Create clustered Index idx_t on t(c1) on ps(c1) -- non-aligned index build

The serial plan is straightforward.

Index Insert (write data to the in-build index)
|
Sort (order by index key)
|
Scan (read data from source)

The sort iterator creates one sort table per target partition (there are four partitions in this example so we will construct four sort tables concurrently). By default, we use the user database for sort to spill data. As we mentioned before, we free each extent from sort table after all its rows are copied. By doing this, for each partition, we can reduce the disk space requirement from 3*partition size (source + sort table + b-tree) to just about 2.2*partition size. Therefore, each file group requires 2.2*(size of all partitions that belong to this file group) of disk space. If the users specify sort_in_tempdb, then all the sort tables reside in the tempdb. Therefore, we require 2.2*(Size of the whole index) of free space in tempdb.

Index insert iterator can start building index after the sort iterator finishes sorting all sort tables. Therefore, we will have as many sort tables as the number of partitions at the same time. Recall that each sort table requires at least 40 pages. So, the minimum required memory will be #PT*40pages.

When it comes to parallel plan, it looks like

X (Exchange)
   |
Index Insert
   |
Sort
   |
Scan

Each worker thread is assigned (Partition Count)/(Degree Of Parallelism) number of partitions (e.g. if we have 4 partitions and 4 worker threads each gets 1 partition), which can be skewed. The sort iterator creates one sort table per assigned partition. Each worker scans the source once and retrieves the rows that belong to its partition(s), the retrieved rows will be inserted into the corresponding sort table depending on which partition they belong to.

After all sort tables got populated, the index builder starts consuming rows from sort tables, it consumes one sort table after another, building b-tree(s) in each partition's file group. Currently workers do not share partitions. Therefore, it is possible for one worker to finish all assigned partitions and idle while another worker is busy inserting rows.

The disk space and memory requirements are exactly the same as the serial plan. This is because in both cases, we cannot start building the index until all the sort tables are populated.

How to Check Whether the Final Query Plan is Optimized for Star Join Queries?

2007-04-09T21:52:00Z

The star join optimization technique is an index based optimization designed for data warehousing scenarios to make optimal use of non-clustered indexes on the huge fact tables. The general idea is to use the non-clustered indexes on the fact table to limit the number of rows scanned from it. More details of index based star join optimization can be found at http://blogs.msdn.com/bi_systems/pages/164502.aspx.

The following discussion is based on SQL Server 2005 query plans.

In SQL Server 2005, we put "StarJoinInfo" element in Showplan XML to highlight star join optimization. If the query plan contains the “StarJoinInfo” element, then SQL Server has identified this plan as a star join plan and it definitely is one.

However, the query optimizer may not detect all star join plans due to star join detection restrictions. Hence there are some star join plans that won’t have the StarJoinInfo. This post will shed some light on how to manually detect if a given query plan is a star join plan.

These steps can help you identify what’s NOT a star join plan:

First identify your fact table.
If you see clustered index scan (or table scan) on fact table, then it’s not an index-based star join plan (however, this is a valid multi-table join plan, which can benefit from multiple-bitmap filter pushdown).

To identify a star join plan, you should:

Again, first identify your fact table
You should have a single RID lookup or clustered index seek on the fact table
Restrictive dimensions (dimension tables with restrictive filters in the query) should be processed before processing the clustered index seek or RID lookup on the fact table. You should find either:

A Cartesian product between the dimensions joined with a multi-column index on the fact table.
Or a semi-join of the dimensions with some non-clustered indexes on the fact table.
Or a join between multiple dimensions.

Non-restrictive dimensions are joined later with the fact table.

So the rule of thumb is: to detect whether a given plan is using index-based star join optimization, you should always look for a seek on a fact table that is based on some joins of some dimension tables.

Hash Warning SQL Profiler Event

2007-02-02T06:18:00Z

One of the less well-known warning events that is logged to SQL Profiler trace is the Hash Warning event. Hash Warning events are fired when a hash recursion or hash bailout has occurred during a hashing operation. Both of these situations are less than desirable, as they mean that a Hash Join or Hash Aggregate has run out of memory and been forced to spill information to disk during query execution. When a hashing operation spills to disk, this almost always results in slower query performance and can cause space consumption increase in tempdb.

Note that the Hash Warning event needs to be explicitly enabled within SQL Profiler. It is not one of the “default” set of events. More info on SQL Profiler can be found here

What can be done if you see a lot of Hash Warning events? The recommended actions are:

· Make sure that statistics exist on the columns that are involved in the hashing operation. Without statistics, the hashing operation has no way to know how much memory to pre-allocate.

· Even if statistics do exist, try updating them, as this can give more current information to the hashing operation when it decides how much memory to allocate.

· Try using a different type of join (this can be done by hinting OPTION(MERGE JOIN) or OPTION(LOOP JOIN)). Note that forcing a join type does not necessarily guarantee a better execution plan.

· If all of these fail, you can increase the available memory on the computer.

A sample of what you will see in the profiler would look something like the following. Note the batch starting, followed by a number of Hash Warning events prior to batch completion. Also, the SPID that is causing the events will be recorded

EventClass	StartTime	SPID
SQL:BatchStarting	2007-02-01 18:53:34.703	51
Hash Warning	2007-02-01 18:53:48.267	51
Hash Warning	2007-02-01 18:53:48.283	51
Hash Warning	2007-02-01 18:53:50.097	51
Hash Warning	2007-02-01 18:54:05.300	51
SQL:BatchCompleted	2007-02-01 18:54:19.130	51

- Steve Herbert

SQL Server Query Execution

Index Build strategy in SQL Server - Part 3: Offline Serial/Parallel Partitioning (Aligned partitioned parallel index build)

2007-01-20T04:20:00Z

Aligned partitioned parallel index build

In case of parallel build we scan and sort partitions in parallel and the actual number of sort tables existing at the same time will depends on the actual number of concurrent workers. Partitions are being chosen by workers one by one and when one worker completes with one partition it takes the next partition which is not yet taken by another worker. Each worker builds 0 - N partitions (we do not share one partition among multiple workers). Why can it be 0? If DOP > # of partitions, then we do not have enough partition to give out to all the workers(s). Which partition a worker is going to work on? This is non-deterministic per execution. First come first serve.

As we never share one partition among several workers, the biggest partition becomes a bottleneck. A situation could happen when all workers completed with their partitions and one worker still sorting the biggest one along. Which also means the resource being used by this query (such as memory and threads) will not be available for other queries.

There is no final stitch among workers for partitioned index. Each partition is being represented as a separated b-tree in storage engine.

How does it affect disc space requirements:

- In case of sorting in user’s database (default setting) the requirements are actually the same as in case of serial build as we are sorting in each filegroup for each corresponding partition we will need 2.2*(Size of partition) for each filegroup.

- In case of using Sort_in_tempdb (SORT_IN_TEMPDB = ON) we won’t have the same advantage as in case of serial build because we may have several sort tables at the same time and as long as we don’t know the actual distribution of data between the partitions we will still require the same 2.2*(Size of the whole index) of free space in tempdb.

Memory consideration:

We will have several sort tables at the same time (depends of #DOP and # of partitions) and we will need at least 40pages per each sort table to be able to start the index build operation. So, the minimum required memory will be #DOP*40pages.

Total memory = 40 * DOP + additional memory.

Note that additional memory does not change for serial or parallel plans, this is because the total number of rows we need to sort remain the same in both plans.