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

Performance Monitoring Tools: Logging with System Monitor

Last updated Mar 28, 2003.

In a previous tutorial I introduced you to a tool that is included with every Windows Server operating system, the Windows System Monitor (also known as Performance Monitor). This tool gives you the ability to track and record changes on various objects on your Windows Workstation or Server. It's a tool used quite often the second phase of the performance tuning process, where you evaluate the system's components.

The previous introduction to the System Monitor focused on navigating and using the tool. This tutorial will show you how to create the output on a schedule, so in effect you can record historical metrics.

Before I get started, I'll revisit how the Windows System Monitor works. This tool uses three primary components to track things on your system.

The first component is System Monitor itself, which displays and controls the output of the things you're monitoring.

The second component is a set of registry entries. System Monitor creates some entries, and others are created by individual applications, such as SQL Server. These registry entries act as placeholders for the values a program can provide.

The last component is the application. Applications write the values into those registry entries that System Monitor created; based on the counters and objects they can track. For instance, SQL Server tracks an object Locks and one of the counters on the Lokcs Objects is is called Average Wait Time (ms). SQL Server writes the time (the value of the counter) of all lock waits into a registry location, and System Monitor reads and displays that value.

Using these three components is very efficient, since none of them will interfere with each other. Using registry entries as the intermediate mechanism is also why you'll see different counters on different systems, since the applications and even service packs change the entries that are there.

In this tutorial, I'll only show a few counters, but Books Online shows many more, based on the processes you are monitoring. What I'm interested in showing you here is the process to track counters, not the specific counters. The important thing is to understand what you're trying to monitor and keep that consistent throughout the monitoring that you do. In other tutorials I'll show you specific counters to track based on what you're trying to locate.

Another factor in choosing the objects and counters that you're after is the impact that the monitoring has on your system. Since the application is architected to create these counter values, there isn't a huge penalty for dealing with them, but the more you monitor, the more work the system has to do in writing and reading the values. For the most part, however, you won't have to be very concerned about the impact of the monitoring. In the example I'll show today, the performance hit the system will take is quite low.

The final factor in your monitoring decision is time. There are two parts to the time dimension of modeling: when you monitor, and the interval you use on the counters.

You should pick a time when there is low or little activity and continue the collection through the heavier periods of use. The monitoring should finish after another period of low activity. This creates a very good set of values to work with when you're using statistical methods of analysis or creating graphs from the data. If you're looking for peak activities, this is less important. It's also not always possible to find downtime, so you may be forced into monitoring only during peak use.

The second time dimension is the granularity of the values, or the interval. If you have a value that peaks every thirty seconds but you're monitoring by five-minute intervals, you'll miss a great deal of the values you're looking for. I have a set of systems at work that I monitor that house a large application, and my interval is five seconds for about 30 counters. I run that for 24 hours, and the log files I create are less than 10 megabytes in size. I pick up those logs and erase them from the server, and there is no noticeable impact to performance in the applications on that server. I do this about once a month.

Now that you understand the generalities of what you want to do, I'll explain how to set up a simple set of counters to be monitored. Before I do that, I'll explain what these counters are for.

My intent is to create a document that shows the pressure a system is under for a sizing exercise. I want to gather a historical set of values that will allow me to show minimum and maximum values, averages, and then standard deviations from those averages.

A standard deviation shows how far the values of a set of numbers are from the average. The closer to 0, the better the chances are that the numbers in the set are really close to the average. For instance, let's assume I have two coins in my pocket. The average value of those coins is 5.5 cents. But the standard deviation of those coins is 6.3 - meaning that I really can't trust that there are a lot of nickels in my pocket. I might actually have a dime and a penny.

The performance values I collect often work out the same way. Perhaps the processor shows that it is very busy for a short time, and then not busy at all. The average would tell me I'm fine at the current speed and number of processors — since it shows only 50% busy — but the standard deviation would show that the average are not be trustworthy. In that case I should pay attention to the distribution of the values through the set, since I want to ensure that the system is fast all of the time. I might have peaks of CPU use that cause the users to complain about the response of the system - if I only use the average of my counters I'll believe that the CPU is fine, when in fact the spikes are the issue.

With all that said, I'll get started. If you want to follow along, I'll show you how. Although the steps I'll use here shouldn't cause any problems on your system, you should always practice on a test server first, just in case. Once you're comfortable with the process, you can use it on your production systems. I'll start with the graphical method of creating a log, and then I'll explain a command-line tool you can use to automate the process.

There are two versions for the graphical tools, one found in earlier Windows Servers and the newer Reliability and Performance Monitor in Windows Server 2008 and higher. I'm focusing on the later versions, but the general concepts are the same.

First, log on to your test server and open the Start menu. In the Administrative Tools menu item you'll find the Reliability and Performance Monitor program. Start that, expand the Monitoring Tools object and then click the Performance Monitor icon and you're shown a screen that contains the Performance Monitor.

The first step in using this graphical tool is to create the objects and counters that you want to track. the Processor object and the % Processor Time counter is already selected, with the _Total set of instances. That means the graph you see is showing the value of the percentage of time the processors (all of them, added as a group because of the _Total setting) is using on the system. It's a general indication of how busy your system is - at least from the processor perspective. You can leave that one running for now, but you want to add a few more.

To do that, click the green plus-sign ( + ) icon  in the top bar of the right-hand pane. That brings up a view similar to this one. 

You'll see a list of objects on your system, such as the SQL Server: Memory Manager I've got here. If you click the Plus-sign next to that object the counters it contains show below it.

At first, all of the counters are selected, until you make a choice. Select the objects you wish, and click the Add >> button to add them to the collection on your screen. Then scroll to other objects and expand those to add more.

Here are a few objects that I am collecting on this test system - add them to yours if you wish.

Object

Counter

Processor

Processor Time (_Total)

SQL Server:Memory Manager

Target Server Memory (KB)

SQL Server:Memory Manager

Total Server Memory (KB)

Logical Disk

Avg. Disk sec/Transfer

Network Interface

Bytes Total/sec (select the network card the users are on)

SQL Server: General Statistics

Transactions

SQL Server: General Statistics

User Connections

With those added, you can see them working the counters across the screen. You can leave the default interval for now (a counter every 1 second) but in production you normally want to set this to a higher value, such as 5 seconds or even higher. The more often you collect these metrics, the larger the logging file and the more impact you'll have on your system.

With the counters added, it's now time to make them into a logging file. Right-click the Performance Monitor icon and select New and then Data Collector Set. That brings up a panel where you can create the settings for the log file - the objects and counters you want to track. They will pick up from the screen selections you just made.

You're first asked to name the collector set - I called mine "SQLServer," with no spaces. That will make it a bit easier for the automation I'll create later. You can call yours anything you like. With that selection made, you can click the Next button.

Next you're asked where to save the collection set. I'm putting mine in the TEMP directory on my C drive, but once again you can place yours wherever you like - just remember to note down the location. Now click Finish.

Your counters will continue to display - but you now have a new object on your system. To display it, and control it, expand the Data Collector Sets icon in the left-hand pane. Expand the User Defined object there and you should see the name you typed in a moment ago.

You can double-click this name if you wish, and from there you can edit the objects and counters it contains. Do that now, and change the format of the file to Comma Separated Values (CSV) from the main panel there. This ensures you can open it in a spreadsheet later - or even import the data into SQL Server if you wish.

You can also click the green arrow icon in the top bar to start the collection of that data into the file you specified - but don't do that just yet. While that will collect the data, odds are you won't want to do this by hand each time you want to collect data. It's far better to make a batch file with the commands I'll show you later. In fact, you can just use the commands in the Windows Task Scheduler or even as a PowerShell line in the SQL Server Agent to schedule the collections to perform baselines or in response to an event that you want to know more about.

Now that you have the Data Collector set, open a command prompt, and optionally navigate to where you saved it. In my case, as you recall, it was on the C:\Temp directory.

From there, simply type the commands you see below, substituting the name you gave your Data Collector Set for mine:

  LOGMAN "SQLServer" Start

Wait a period of time - in my case I waited just a few minutes - and then type (or run from an automation tool as I described) the command:

  LOGMAN "SQLServer" Stop

If you list the contents of that directory you'll see the CSV file you made of the counters. From there it's a simple matter of importing the CSV file into your spreadsheet, or even use the BULK IMPORT command in Transact-SQL or even SQL Server Integration Services (SSIS) to import the file into a table.

I use Microsoft Office Excel quite often, so I'll import mine into that. I open the directory in Windows, then right-click the file and open it in Excel. I changed the format of the first column to time, and the headings of the second and third columns to something more easy to read - on my system I just collected two counters for this display. Then I made a graph of the results.

In PowerShell, you can even automate the creation of the spreadsheet, create a web page instead, send the data in an e-mail, or even perform calculations on the data if you wish. The possibilities are wide open.

I automate this type of collection at least once a month, and then do comparisons from month to month to be able to predict when I need to add hardware, more aggressively tune systems, or even when I should roll one system over to another. I also use it to see which systems can be consolidated or virtualized.