Home > Articles > Data > SQL Server

SQL Server Reference Guide

Hosted by

Toggle Open Guide Table of ContentsGuide Contents

Close Table of ContentsGuide Contents

Close Table of Contents

General Transact-SQL (T-SQL) Performance Tuning, Part 2

Last updated Mar 28, 2003.

I’m in a series of tutorials that explains how to tune the queries you have within your system. If you haven’t read the first tutorial, it’s a good idea to stop here and take care of that now. Keep in mind as you read through this series that tuning the queries in your system is only part of the picture. There are other components for tuning the entire system, and I’ve covered that process in this section of the Guide.

But I will say this — no matter how much you tune the network, hardware, server settings and even the high-level code in the client application, nothing can save time more than tuning the queries that the application runs. You see, it’s the work you avoid that goes fastest — but I’m getting ahead of myself a bit.

In the last installment of this series, I explained that “knowledge is power.” This means that you need to set up your environment so that you can observe the queries in a “clean-room” where you know that other activities won’t affect the queries you’re examining. That’s the first part of any scientific endeavor — to ensure that you have control over the experiment.

In this tutorial, I’ll explain the second part of the scientific experiment, which is observation. I’ll explain the primary tool that you can use to examine your queries as they run on the server, and show you how to interpret the results. In the final installment I’ll explain a few tips for speeding up the queries by avoiding the problems that you’ll find with the knowledge and the tools that I show you here.

So let’s get started.

Query Plans

As I explained in the last tutorial, a query isn’t handled in the same order that you send it. The SQL Server engine parses the query and then develops the fastest way to answer it. This path to the data is called a “Query Plan,” and you can access it in multiple outputs, as I’ll show you in a moment.

Once again, I’ll emphasize that you should do all of your testing on a separate system from production. You can actually slow down your queries in production, so don’t do any of this work there. You should also download and install the sample databases for SQL Server, as I’ve explained in previous tutorials. That will allow you to work through the examples I have here, and then once you’re comfortable you can test out the concepts on your production code.

I explained that the query moves through several layers between the time it is called on the client and the time the data is returned to the caller. What I’m focusing here are only the actions performed in the Engine, not along any of those other paths.

When the Query Plan is created (called “compiled”), it is stored up in memory so that in case another query asks for the same data, it has the fastest path already available. This happens more often than you think — if the tables are normalized to the proper level, queries are often looking for the same data.

As time goes by, if a Query Plan isn’t accessed as often as other Plans are, it will be “aged out,” or dropped from memory. Also, if the path to the data changes dramatically the plan will be “re-compiled,” even if the data requests are the same.

With that background, we can begin to examine the plans, and to learn how to monitor them for compiles and recompiles.

Viewing the Graphical Query Plan Output

There are two modes for displaying the Query Plan graphically. The first is to display the “estimated” Query Plan. In this mode, the system doesn’t actually run the query, it just tries to figure out what it would do if it actually did run it. This can be very useful when you’re running against a large set of data, and you’d like to take a look at the plan first.

You can try this out yourself in SQL Server Management Studio in SQL Server 2005 and higher. Copy the query from below, and make sure you’re in the “AdventureWorks” sample database:

FROM Sales.SalesOrderDetail 
WHERE UnitPrice < 25.00
HAVING AVG(OrderQty) > 5

Now “highlight” that query by selecting it with your mouse. Then press CTRL-L, or select “Display estimated execution plan” from the “Query” menu.

Another tab appears when you do this, and you’re presented with several graphics and lines, moving from right to left. Hold on a moment and I’ll explain what you should know about those.

The “estimated” Query Plan has its uses, but unless the query is quite long or returns a huge amount of data, I recommend that you use the “actual” Query Plan. To do that, highlight that query again and press CTRL-M, or select “Include Actual Execution Plan” from the “Query” menu.

When you run the query, the results are returned as expected, but now you’ll see another tab with the Execution Plan in it. By the way, an “Execution Plan” and a “Query Plan” are basically the same thing.

There’s quite a bit going on here, so take a moment and look at the screen. First, just at the top you’ll notice that there is a statement telling you that the “query cost to the batch” is 100%. That’s because you only have one query in this batch. If you sent two statements, you would see two plans here with a relative cost. What you’re looking to do here is to go after the query that takes the longest, first.

Using that same logic, you can look down in the graphical area to see the operations that the system took to get to the data, and how much work each operation took. If you take your mouse pointer and leave it over one of these icons you’ll see a lot of information about what the query was doing at the time. Underneath each icon is the amount of work as a percentage each operation took.

In my case, the first icon on the far right represents a “Clustered Index Scan.” This means that the system used a Clustered Index (more on those here) to find the data, but it had to read the entire Index to make sure it got to all the data. I would rather see a “lookup” operation here, since that would find all of the data I need faster. I’ll explain a few more of these icons shortly, and then tell you where you can look them all up.

Moving to the left, you see a large arrow. If you hover over an arrow with the mouse, you’ll see information on the amount of rows the query sends to the next operation. If you look closely you’ll see where the arrows are “fatter” or bulkier and which are “skinny” or smaller. Thicker lines mean more rows, thinner ones mean fewer.

There are quite a few operations shown in this graphic. The ones that should stand out immediately are those that indicate any kind of a “Scan” operation, unless there is less than 1,000 rows involved. It’s quicker to read 1,000 rows that to do anything else, so the Engine will almost always pick that mechanism in that case. A Scan operation means that the Engine read every single row in a table or index to make sure it gets all of the data. If you see a Scan, you might want to consider an Index — if, of course, that doesn’t affect the INSERT or UPDATE operations in a bad way. More on that in the Index tutorial I have here.

You can see all of the graphical icons here. I pay attention to the Scan operations, as I mentioned, and to the WHILE loops, and also the arrows that show branching and merging. I’ll explain more about these later, but if you review the last link you can read a little ahead.

You can save these plans out to the hard drive by right-clicking inside the plan and selecting “Export” from the menu that appears. The interesting thing is that you don’t have to have the save query on another system to open it, or even the same database — the plan is something separate altogether. Anyone else with the same version of SQL Server Management Studio (SSMS) can open it.

Viewing the Text and XML Query Plan Output

You can also view the Query Plan in text and XML formats. This is useful when you don’t have access to the graphical tools, or you want to save the output in a format to work with later or analyze with other tools.

You accomplish this with the “SET” commands. The first is called SET SHOWPLAN_TEXT ON. When you issue this statement, your query produces a text representation of that same graphical information you saw earlier. SET SHOWPLAN_TEXT ON doesn’t have quite as much information, however, but you can fix that with another command, SET SHOWPLAN_ALL ON.

To see the data in XML, use the command SET SHOWPLAN_XML ON. In all these cases, you need to use the SET SHOWPLAN_TEXT OFF, or SET SHOWPLAN_ALL OFF, or the SET SHOWPLAN_XML OFF statements to display the normal results of your query.

Those are the basics of displaying the Query Plans — but I haven’t explained what to do with the information yet. I’ll do that in future tutorials, but even having all this information doesn’t form the full picture of what the query is doing to the system. To get the full impact of the query on the system, you need to combine what the query is doing with a view of the Meta Data about how the system deals with those impacts.

I’ve covered a lot of the views of the data in the Query Plans, and the factors that affect them, in the series of articles on Meta-Data. You can use the combination of those views and the information in the Meta Data views to zero in on the problems in your queries — which is what I’ll cover in the next article in this series.

InformIT Articles and Sample Chapters

If you’ve come into this tutorial without reading the complete series I have on performance tuning, you can find that here.

Books and eBooks

You can’t do better than SQL Performance Tuning by Peter Gulutzan and Trudy Pelzer.

Online Resources

Here’s the best SQL Server Performance Tuning site on the planet.