After a few months, we're pretty happy with the content and readership we have created.  This is a great way to work more closely with our customers and to give them guidance that helps them be effective using our software.

One thing we have realized, however, is that we'd also like to get our query execution team involved in this effort as well.  So, we're creating a new blog that will cover both teams.  The focus of the blog will remain the same but we'll get additional perspectives from some of our experts on issues like memory tuning, prefetching, locking, that we haven't had as much time to cover so far.

Please join us over at the new blog:

http://blogs.msdn.com/sqlqueryprocessing/default.aspx

Thanks,

Conor

(2006-09-01 added a paragraph on parallel query plans)

 

In SQL Server, “Statistics Profile” is a mode in which a query is run where you can see the number of invocations of each physical plan operator in the tree.  Instead of running a query and just printing the output rows, this mode also collects and returns per-operator row counts.  Statistics Profile is used by the SQL Query Optimizer Team to identify issues with a plan which can cause the plan to perform poorly.  For example, it can help identify a poor index choice or poor join order in a plan.  Oftentimes, it can help identify the needed solution, such as updating statistics (as in the histograms and other statistical information used during plan generation) or perhaps adding a plan hint.  This document describes how to read the statistics profile information from a query plan so that you can also debug plan issues.

 

A simple example query demonstrates how to retrieve the statistics profile output from a query:

 

use nwind

set statistics profile on

select * from customers c inner join orders o on c.customerid = o.customerid;

 

The profile output has a number of columns and is a bit tricky to print in a regular document.  The key pieces of information that it prints are the plan, which looks like this:

 

StmtText                                                                                               

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

select * from customers c inner join orders o on c.customerid = o.customerid                          

  |--Hash Match(Inner Join, HASH:([c].[CustomerID])=([o].[CustomerID]),

       |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Customers].[aaaaa_PrimaryKey] AS [c]))          

       |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Orders].[aaaaa_PrimaryKey] AS [o]))             

 

Other pieces of useful information are the estimated row count and the actual row count for each operator and the estimated and actual number of invocations of this operator.  Note that the actual rows and # of executions are physically listed as early columns, while the other columns are listed later in the output column list (so you typically have to scroll over to see them).

 

Rows                 Executes           

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

1078                 1                  

1078                 1                  

91                   1                  

1078                 1                  

 

EstimateRows            

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

1051.5834               

1051.5834               

91.0                    

1078.0                  

 

EstimateExecutions      

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

NULL

1.0

1.0

1.0

 

Other fields, such as the estimated cost of the subtree, the output columns, the average row size, also exist in the output (but are omitted for space in this document).

 

Note: The output from statistics profile is typically easiest to read if you set the output from your client (Query Analyzer or SQL Server Management Studio) to output to text, using a fixed-width font.  You can then see the columns pretty easily and you can even move the results into a tool like Excel if you want to cluster the estimates and actual results near each other by rearranging the columns (SSMS also can let you do this).

 

There are a few key pieces of information needed to understand the output from Statistics profile.  First, query plans are generally represented as trees.  In the output, children are printed below their parents and are indented:

StmtText                                                                                               

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

select * from customers c inner join orders o on c.customerid = o.customerid                           

  |--Hash Match(Inner Join, HASH:([c].[CustomerID])=([o].[CustomerID]),

       |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Customers].[aaaaa_PrimaryKey] AS [c]))          

       |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Orders].[aaaaa_PrimaryKey] AS [o]))             

 

In this example, the scan of Customers and Orders are both below a hash join operator.  Next, the “first” child of the operator is listed first.  So, the Customers Scan is the first child of the hash join.  Subsequent operators follow, in order.  Finally, query execution plans are executed in this “first” to “last” order.  So, for this plan, the very first row is returned from the Customers table.  (A more detailed discussion of operators will happen later in the article).  Notice that the output from the graphical showplan is similar but slightly different:

 

(see attachment tree.jpg).

 

In this tree representation, both Scans are printed to the right on the screen, and the first child is above the other operators.  In most Computer Science classes, trees are printed in a manner transposed from this:

 

     Hash Join

     /       \

 Customers  Orders

 

The transposition makes printing trees in text easier.  In this classical view, the tree is evaluated left to right with rows flowing bottom to top. 

 

With an understanding of the nuances of the query plan display, it is possible to understand what happens during the execution of this query plan.  The query returns 1078 rows.  Not coincidentally, there are also 1078 orders in this database.  Since there’s a Foreign Key relationship between Orders and Customers, it requires that a match exist for each order to each customer.  So, the 91 rows in Customers match the 1078 rows in Orders to return the result.

 

The query estimates that the join will return 1051.5834 rows.  First, this is a bit less than the actual (1078) but is not a substantial difference.  Given that the Query Optimizer is making educated guesses based on sampled statistical information that may itself be out-of-date, this estimate is actually pretty good.  Second, the number is not an integer because we use floating point for our estimates to improve accuracy on estimates we make.  For this query, the number of executions is 1 for both the estimate and actual.  This won’t always be the case, but it happens to be true for this query because of the way hash joins work.  In a hash join, the first child is scanned and a hash table is built.  Once the hash join is built, the second child is then scanned and each row probes the hash table to see if there is a matching row. 

 

Loops join does not work this way, as we’ll see in a slightly modified example. 

 

select * from customers c with (index=1) inner loop join orders o with (index=1) on c.customerid = o.customerid

 

In this example, I’ve forced a loop join and the use of clustered indexes for each table.  The plan now looks like this:

 

StmtText                                                        

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

select * from customers c with (index=1) inner loop join orders o

  |--Nested Loops(Inner Join, WHERE:([nwind].[dbo].[Orders].[Cust

       |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Customers].

       |--Table Spool                                           

            |--Clustered Index Scan(OBJECT:([nwind].[dbo].[Orders

 

Beyond the different join algorithm, you’ll notice that there is now a table spool added to the plan.  The spool is on the second child (also called the “inner” child for loops join because it is usually invoked multiple times).  The spool scans rows from its child and can store them for future invocations of the inner child.  The actual row count and execution count from the statistics profile is a bit different from the previous plan:

 

Rows                 Executes             

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

1078                 1                   

1078                 1       <--- Loop Join

91                   1       <--- Scan of Customers

98098                91      <--- Spool

1078                 1       <--- Scan of Orders

 

In this plan, the second child of the loop join is scanned 91 times returning a total number of 98098 rows.  For the actual executions, the total number of rows is to sum of all invocations of that operator, so it is 91*1078=98098.  This means that the inner side of this tree is scanned 91 times.  Nested Loops joins require rescans of the inner subtree (Hash Joins do not, as you saw in the first example).  Note that the spool causes only one scan of the Orders table, and it only has one execution as a result.  It isn’t hard to see that there are far more rows touched in this plan compared to the hash join, and thus it shouldn’t be a huge surprise that this plan runs more slowly.

 

Note: When comparing the estimated vs. actual number of rows, it is important to remember that the actual counts need to be divded by the actual number of executions to get a value that is comparable to the estimated number of rows returned.  The estimate is the per-invocation estimate.

 

As a more complicated example, we can try something with a few more operators and see how things work on one of the TPC-H benchmark queries (Query 8, for those who are interested, on a small-scale 100MB database):

 

SELECT O_YEAR,

       SUM(CASE      WHEN   NATION = 'MOROCCO'

                     THEN   VOLUME

                     ELSE   0

                     END) / SUM(VOLUME)   AS MKT_SHARE

FROM   (      SELECT datepart(yy,O_ORDERDATE)          AS O_YEAR,

                     L_EXTENDEDPRICE * (1-L_DISCOUNT)  AS VOLUME,

                     N2.N_NAME                         AS NATION

              FROM   PART,

                     SUPPLIER,

                     LINEITEM,

                     ORDERS,

                     CUSTOMER,

                     NATION N1,

                     NATION N2,

                     REGION

              WHERE  P_PARTKEY     = L_PARTKEY AND

                     S_SUPPKEY     = L_SUPPKEY AND

                     L_ORDERKEY    = O_ORDERKEY AND

                     O_CUSTKEY     = C_CUSTKEY AND

                     C_NATIONKEY   = N1.N_NATIONKEY AND

                     N1.N_REGIONKEY       = R_REGIONKEY AND

                     R_NAME        = 'AFRICA' AND

                     S_NATIONKEY   = N2.N_NATIONKEY AND

                     O_ORDERDATE   BETWEEN '1995-01-01' AND '1996-12-31' AND

                     P_TYPE        = 'PROMO BURNISHED NICKEL' AND

                     L_SHIPDATE    >= CONVERT(datetime,(1156)*(30),121) AND                      L_SHIPDATE       <  CONVERT(datetime,((1185)+(1))*(30),121)     

       )      AS     ALL_NATIONS

GROUP  BY     O_YEAR

ORDER  BY     O_YEAR

 

As the queries get more complex, it gets harder to print them in a standard page of text.  So, I’ve truncated the plan somewhat in this example.  Notice that the same tree format still exists, and the main operators in this query are Scans, Seeks, Hash Joins, Stream Aggregates, a Sort, and a Loop Join.  I’ve also included the actual number of rows and actual number of executions columns as well.

 

Rows  Executes   Plan

0      0 Compute Scalar(DEFINE:([Expr1028]=[Expr1026]/[Expr1027]))             

2      1   |--Stream Aggregate(GROUP BY:([Expr1024]) DEFINE:([Expr1026]=SUM([par

2      1        |--Nested Loops(Inner Join, OUTER REFERENCES:([N1].[N_REGIONKEY]

10     1             |--Stream Aggregate(GROUP BY:([Expr1024], [N1].[N_REGIONKEY

1160   1             |    |--Sort(ORDER BY:([Expr1024] ASC, [N1].[N_REGIONKEY] A

1160   1             |         |--Hash Match(Inner Join, HASH:([N2].[N_NATIONKEY

25     1             |              |--Clustered Index Scan(OBJECT:([tpch100M].[

1160   1             |              |--Hash Match(Inner Join, HASH:([N1].[N_NATI

25     1             |                   |--Index Scan(OBJECT:([tpch100M].[dbo].

1160   1             |                   |--Hash Match(Inner Join, HASH:([tpch10

1000   1             |                        |--Index Scan(OBJECT:([tpch100M].[

1160   1             |                        |--Hash Match(Inner Join, HASH:([t

1160   1             |                             |--Hash Match(Inner Join, HAS

1432   1             |                             |    |--Hash Match(Inner Join

126    1             |                             |    |    |--Clustered Index

0      0             |                             |    |    |--Compute Scalar(D

224618 1             |                             |    |         |--Clustered I

0      0             |                             |    |--Compute Scalar(DEFINE

45624  1             |                             |         |--Clustered Index

15000  1             |                             |--Index Scan(OBJECT:([tpch10

2      10            |--Clustered Index Seek(OBJECT:([tpch100M].[dbo].[REGION].[

 

I’ll point out a few details about the statistics profile output.  Notice that the Compute Scalars (also called Projects) return zero for both columns.  Since Compute Scalar always returns exactly as many rows as it is given from its child, there isn’t any logic to count rows again in this operator simply for performance reasons.  The zeros can be safely ignored, and the values for its child can be used instead.  Another interesting detail can be seen in the last operator in this printout (the Seek into the Region table).  In this operator, there are 10 executions but only 2 rows returned.  This means that even though there were 10 attempts to find rows in this index, only two rows were ever found.  The parent operator (the Nested Loops near the top) has 10 rows coming from its first (left) child and only 2 rows output by the operator, which matches what you see in the seek.  Another interesting tidbit can be found if you look at the estimates for the Seek operator:

 

Est.# rows  Est. #executes

1.0                                           20.106487

 

The SQL Server Query Optimizer will estimate a minimum of one row coming out of a seek operator.  This is done to avoid the case when a very expensive subtree is picked due to an cardinality underestimation.  If the subtree is estimated to return zero rows, many plans cost about the same and there can be errors in plan selection as a result.  So, you’ll notice that the estimation is “high” for this case, and some errors could result.  You also might notice that we estimate 20 executions of this branch instead of the actual 10.  However, given the number of joins that have been evaluated before this operator, being off by a factor of 2 (10 rows) isn’t considered to be too bad.  (Errors can increase exponentially with the number of joins).

 

SQL Server supports executing query plans in parallel.  Parallelism can add complexity to the statistics profile output as there are different kinds of parallelism that have different impacts on the counters for each operator.  Parallel Scans exist at the leaves of the tree, and these will count all rows from the table into each thread even through each thread only returns a fraction of the rows.  The number of executions (the second column in the output) will also have 1 execution for each thread.  So, it is typical to just divide the number of threads into the total number of rows to see how many rows were actually returned by the table.  Parallel zones higher in the tree usually work the same way.  These will have N (where N is the degree of parallelism) more executions than the equivalent non-parallel query.  There are a few cases where we will broadcast one row to multiple threads.  If you examine the type of the parallelism exchange operation, you can identify these cases and notice that one row becomes multiple rows through the counts in the statistics profile results.

 

The most common use of the statistics profile output is to identify areas where the Optimizer may be seeing and using incomplete or incorrect information.  This is often the root cause of many performance problems in queries.  If you can identify areas where the estimated and actual cardinality values are far apart, then you likely have found a reason why the Optimizer is not returning the “best” plan.  The reasons for the estimate being incorrect can vary, but it can include missing or out-of-date statistics, too low of a sample rate on those statistics, correlations between data columns, or use of operators outside of the optimizer’s statistical model, to name a few common cases.

 

Future posts will cover some of the details associated with tracking down each of these cases.

If you read the Books Online page describing the UPDATE STATISTICS command, you will see that there are some undocumented options.

 

UPDATE STATISTICS table | view
    [
        {
            { index | statistics_name }
          | ( { index |statistics_name } [ ,...n ] )
                }
    ]
    [    WITH
        [
            [ FULLSCAN ]
            | SAMPLE number { PERCENT | ROWS } ]
            | RESAMPLE
            | <update_stats_stream_option> [ ,...n ]
        ]
        [ [ , ] [ ALL | COLUMNS | INDEX ]
        [ [ , ] NORECOMPUTE ]
    ] ;

<update_stats_stream_option> ::=
    [ STATS_STREAM = stats_stream ]
    [ ROWCOUNT = numeric_constant ]
    [ PAGECOUNT = numeric contant ]

 

<update_stats_stream_option>

    This syntax is for internal use only and is not supported. Microsoft reserves the right to change this syntax at any time.

 

There is a very good reason why these options are undocumented. They are meant for testing and debugging purposes, and should never ever be used on production systems.

 

However, these options can also be extremely helpful to create examples and sample scripts to demonstrate various features and behaviors of the Query Optimizer and the related query plan shapes. We decided to explain what ROWCOUNT and PAGECOUNT do, so that to be able to use these commands in examples we’ll be posting in some future. Feel free to use these options on development systems for experimental and educational purposes, but please do not play with fire and do not use them in production!

 

As the name of these options suggest, ROWCOUNT and PAGECOUNT alter the internal metadata of the specified table or index by overriding the counters containing the row and page counts of the object. These counters are in turn read by the Query Optimizer when processing queries that access the table and/or index in question. These commands can basically cheat the Optimizer into thinking that a table or index is extremely large.

 

SQL Server’s Query Optimization process is structured in multiple stages. Further optimization stages consider a progressively larger and more sophisticated set of possible tree transformations and query optimizations. Later stages of optimization are only entered when the estimated cost of the query is sufficiently high, in order to avoid wasting precious CPU cycles against simple queries that do not need that level of sophistication anyway. The multiple optimization stages are a mean to produce efficient query plans without consuming excessive amounts of CPU. Typically, in order to make the Optimizer “think” a lot and enter these later stages it is necessary to have big tables with a large number of rows and pages, which in turn take time and space to populate. Using ROWCOUNT and PAGECOUNT allows us to exercise these code paths with relatively simple scripts that do not require an extremely complex setup phase.

 

Here is an example. When running this simple script on your SQL 2005 instance you will likely see a different query plan for the two selects before and after updating the statistics. The recompile option is used to ensure that the query plans are regenerated. From the statistics profile, you’ll also see very different estimated row counts and consequently costs.

 

use tempdb

go

 

create table t1(i int, j int)

go

 

create table t2(h int, k int)

go

 

set statistics profile on

go

 

select distinct(i) from t1

go

 

select * from t1, t2 where i = k order by j + k

go

 

update statistics t1 with rowcount = 10000, pagecount = 10000

update statistics t2 with rowcount = 100000, pagecount = 100000

go

 

select distinct(i) from t1 option (recompile)

go

 

select * from t1, t2 where i = k order by j + k option (recompile)

go

 

Enjoy!

Stefano

A question we are frequently asked is what happens when an update statement assigns a column to its same current value. For example,

 

use tempdb

go

 

create table t(i int, cc as i + 1)

create index t_cc on t(cc)

go

 

insert into t values(1)

go

 

update t set i = 1

go

 

update t set i = i

go

 

Columns do get updated even if the value does not change. However, to be honest, i wouldn't worry too much about it.
The runtime cost for updating a row is roughly equal to

a) locating the row in the heap or B-Tree

b) locking the row

c) changing the updated column values

d) logging the change.

For updates to columns that fit in the 8Kb page, i don't think that avoiding part of the copy would make too much of a difference. Realistically, performing a comparison to tell what changed with memcmp would roughly cost as much as changing the value with memcpy. And this cost would surely be overshadowed by a) + b) + d). Updates to columns that don't fit in the 8Kb page are inherently slower, and i think it is up to the application to avoid making unnecessary changes to these.

Now a), b), c) and d) need to be performed for every index carrying one or more columns being modified. So a major factor for the runtime cost of an update query is the number of nonclustered indexes that need to be maintained. In SQL 2005 we introduced an optimization to skip nonclustered index maintenance if none of the updated columns stored in the index actually changed. The reasoning is that running the comparison for only the modified columns stored in one or more nonclustered indexes introduces a runtime overhead that is very small, and surely significantly lower than a), b) and d). Given that columns stored in an index are typically pretty small in size, it is very easy to break even if the optimization actually saves a small minority of nonclustered index row updates. The optimization does not apply to the clustered index, as we want to ensure to always exclusively lock the affected rows even if no columns are really changing.

You can see the new optimization in action by comparing the statistics profile output for one of the update statements in the previous example in SQL 2000 and 2005 (see attached image).

set statistics profile on
go

update t set i = i

go

A question came in about when to use the FAST hint.

If you remember from the row goals post (http://blogs.msdn.com/controlpanel/blogs/posteditor.aspx?SelectedNavItem=Posts&sectionid=6255&postid=564912), it is possible to get a different plan based on the row goal (for example, you can get a nested loops plan instead of a merge/hash join if you have a low enough row goal).  TOP is a common reason to get a row goal in the optimization process.  TOP queries can also be used to page through a result set using a series of queries (if you remember the last row you saw + a few other conditions).

The FAST hint is a way to get the plan you would get with TOP but without the hard limit on the number of rows being returned.  So, if you want to get one "page" of rows quickly, you can use FAST to enable that.  If you want the ability to continue reading from the result set, you can use the FAST hint to get that "give me rows quickly" plan. 

Typically, ISVs that do paging themselves without server-side cursors may consider this hint.  In general, I wouldn't recommend that you use it unless you have a good reason to do so.  If you can ignore it, as with all hints, please do so.  If you have really specific requirements on how your application should work, then you can consider it as a technique that picks more loops joins and minimizes stop-and-go operators, where possible.

Thanks,

Conor

Today's topic is a general primer on something that is called "local-global aggregation".  Effectively, this is a technique to allow you to take a query containing joins and group bys and perform the group by partially "before" the joins.  This can dramatically reduce the number of rows that a query has to process to return a result, and therefore it is a very powerful operation.  This technique is also quite useful to split a grouping operation into something that can be done in parallel.  That obviously can also speed your query as well.

I hope to put something more formal together soon, work permitting, but here is an introduction that you can read in the meantime:

http://citeseer.ist.psu.edu/jaedicke97framework.html

Thanks,

Conor

The Optimizer model makes several assumptions when making plan choice decisions.  These decisions can be false for particular queries and data sets, and sometimes this can cause plan issues.  One such problem is called “The Sorted Seek Problem” by the Optimizer Team, and it affects plan selection, usually between clustered index or heap scan plans and bookmark lookup plans.

 

When costing a series of bookmark lookup operations (for a heap) or index seek operations (for an index), the costing code will effectively treat these as a series of random I/Os.  These are much more expensive than sequential I/Os because the disk arm needs to be moved and the disk needs to spin around to find the right sector(s).  When the correct page is found, it is loaded into the buffer pool and the engine will cache it for future references.  The costing code also understands, on average, how often a page is likely to be found on subsequent operations in the cache, and it costs these accesses much more cheaply than the original access.

 

When a query needs to perform lookups into an access method, the costing code makes the assumption that rows are uniformly distributed over the pages in the table/index.  Under this assumption, effectively each row that is accessed is expensive because we assume that it needs to be loaded from disk.  (At some point, all pages may be in the buffer pool and no longer require the expensive random I/O operation to load a different record from a previously accessed page).

 

In practice, rows are usually not randomly distributed.  In fact, many common user patterns will insert rows in physically clustered order.  For example, you likely would insert all of the order details for an order at the time that the order was created.  These could end up clustered on a very few data pages.  As such, a plan that does a bookmark or index lookup into this structure may complete more quickly than the costing model has forecast.  As a result, the Query Optimizer may pick a different plan that runs more slowly for the customer’s installation. 

 

To be fair, there are a number of scenarios where the rows are uniformly distributed, so they would be broken under a different assumption.  In the future, we may be able to provide additional features to find and fix such problems for customers.  Today, your best option in such situations is to use query hints to force the index plan if you know that you have data that is physically clustered.

 

The following sanitized customer example demonstrates this scenario in more detail.  The table is contains about 80 columns and 5.5 million rows.  In the query, the filter qualifies just over 21,000 rows. 

 

SELECT * FROM CustTable  WHERE col='someconstvalue'

 

  |--Parallelism(Gather Streams)

       |--Table Scan(OBJECT:(CustTable), WHERE:(col=’someconstvalue’))

 

Table ‘CustTable’. Scan count 3, logical reads 416967, physical reads 22, read-ahead reads 413582, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Completed in 3 minutes 9 seconds

 

SELECT * FROM CustTable with (index=idxoncol) WHERE col='someconstvalue'

  |--Nested Loops(Inner Join, OUTER REFERENCES:([Bmk1000], [Expr1005]) WITH UNORDERED PREFETCH)

       |--Index Seek(OBJECT:(idxoncol), SEEK:(col=’someconstvalue’) ORDERED FORWARD)

       |--RID Lookup(OBJECT:(CustTable), SEEK:([Bmk1000]=[Bmk1000]) LOOKUP ORDERED FORWARD)

 

Table 'CustTable'. Scan count 1, logical reads 21167, physical reads 387, read-ahead reads 1354, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Completed in 11 seconds

 

The first query is the original user query, and the default plan picked is a parallel scan of the base table, applying the filter as a pushed non-SARGable predicate into the scan.  It takes over 3 minutes to complete. The second plan uses an index hint to force a bookmark lookup plan.  It completes in 11 seconds.  If you assume that a random disk seek takes approximately 10 ms, you can’t do more than 100 in a second.  For this query, 21,000 seeks would probably take over 3 minutes to do if each was a random I/O against a single disk.  Therefore, there are likely fewer than 21,000 random I/Os happening.  We can see this from the statistics I/O output – there are only 387 + 1354 actual page reads in the query.

 

The actual data distribution is not uniform.  In fact, it is heavily physically grouped.  I ran a query that counted the physical row position in the heap and then filtered out the rows that qualified in this query.  The rows should be somewhat randomly distributed through the set of rows to match the uniformity assumption.  I copied this output into Excel and charted it:

 

[see attachment] 

 

(The pre-filter row number is the y-axis, and the post-filter row number is the x-axis).

 

As you can see, the rows are very clustered on the disk.  As such, many of the rows used in the bookmark lookups are found on pages that are already loaded into the buffer pool from previous I/Os.  This significantly speeds up the query. 

 

The recomendation for these cases is to use query hints if this is a significant issue for your query.

 

Enjoy,

 

Conor

I saw a post in one of the newsgroups today that referenced a piece of information Microsoft published on how the Optimizer makes decisions about remoting certain operations.

http://msdn.microsoft.com/library/default.asp?url=/library/en-us/optimsql/odp_tun_1a_6oxf.asp

It's slightly out-of-date, and I'll work on trying to get it updated.  Specifically, SQL 2005 will remote uniqueidentifiers (and it will support statistics over them as well).  We'll remote queries from SQL 2005 to SQL 2000 that contain filters comparing columns to constant uniqueidentifier values as well.

We published a research paper last year on how the Distributed Query feature works in more detail.  While it does not cover every practical detail of the implementation, you may find it as an interesting reference if you use distributed queries.

http://citeseer.ist.psu.edu/732761.html

If you have other remoting questions/problems, please post comments on them and we'll see if we can get them answered for you.

Thanks,

Conor