Thanks, Guy.

I recently answered this customer question and was looking forward to post the solution here. So here you go!

There is already a lot of information about how Soft NUMA works and how to configure Soft NUMA on your system. So without delving into the basics, I am jumping right ahead to the topic.

Soft Affinity Setting

SQL Server 2000 and above supports the affinity option for greater than 32 processors (33 to 64) only on 64-bit operating system.

If you are running SQL Server 2000 or above on a system with greater than 32 processors, then you can use sp_configure ‘affinity mask’ (32 bit mask) option for the first 32 CPUs and sp_configure ‘affinity64 mask’ (32 bit mask) option for the CPUs 33 to 64.

Ex: Consider a machine with 33 CPUs. Setting the ‘affinity mask’ to 0x0000ffff and ‘affinity64 mask’ to 0x1 hard affinitizes CPU 1 to 16 and CPU 33 to ONLINE.

sp_configure 'show advanced options', 1; RECONFIGURE; GO sp_configure 'affinity mask', 65535; RECONFIGURE; GO sp_configure ‘affinity64 mask’, 1; RECONFIGURE; GO

Registry Settings  

The other day I came across a nice article written by Mario Broodbakker called "SQL Server Wait Events: Taking the Guesswork out of Performance Profiling" (see references) on how to troubleshoot slow response times in SQL Server 2000 and 2005 using the YAPP Method. According to the author, one of the biggest challenges to apply this method in SQL Server is that fact that the wait statistics are collected and exposed at the server level (server wide):

... "This is simply not enough, unless you have the luxury of testing your application and SQL statements on an isolated server" ..." But, of course, it's often hard or impossible to isolate and replicate production problems on an isolated test server." ...

I decided to write this because in SQL Server 2008 the wait statistics are available at virtually any level, and (at the time of writing) I couldn't find any articles on the new sqlos.wait_info and sqlos.wait_info_external extended events.

YAPP Method

The YAPP Method: "Yet Another Performance Profiling Method" is based on a paper written by Anjo Kolk (see references). The basic idea of the method can be expressed with the following formula:

(1)   Response time = service time + wait time

This method for performance troubleshooting has been around for some time and it is a fairly standard practice when debugging slow response times.

You can get the response time and the service time of a query by looking at sys.dm_exec_sessions, columns total_elapsed_time and cpu_time. Alternatively, you can use SQL trace or the more lightweight extended events in SQL Server 2008. Knowing response time and service time and applying formula (1), gives you wait time. So, we know how much we waited. However, in order to get the individual wait types we need the new extended events.

Per session wait statistics in SQL Server 2008

In SQL Server 2008 you can get wait statistics for a given session or statement by leveraging the new extended events. Optionally, you can limit the collection to a small set of wait types or collect only waits exceeding certain duration. If your event session is written carefully, you can use this technique on production systems with very little or no performance impact. In fact, SQL Server 2008 ships with a built-in event session, which is collecting wait statistics, and is running by default out of the box.

Here is how to create an event session to collect wait types for session 51. The events will go to the specified trace file:

create event session session_waits on server

add event sqlos.wait_info

(WHERE sqlserver.session_id=51 and duration>0)

, add event sqlos.wait_info_external

(WHERE sqlserver.session_id=51 and duration>0)

add target package0.asynchronous_file_target

      (SET filename=N'c:\wait_stats.xel', metadatafile=N'c:\wait_stats.xem');

Here is how to start the event session, run your workload, and then stop the event session:

alter event session session_waits on server state = start;

exec sp_workload;

alter event session session_waits on server state = stop;

Here is how to read the trace file:

select * from from sys.fn_xe_file_target_read_file

      ('c:\wait_stats*.xel', 'c:\wait_stats*.xem', null, null)

The result is a list of wait events, representing individual waits, having wait type and duration column. In order to make better sense out of this we need to group the events by wait type and calculate the total duration per wait type. The following code can do this:

create view dbo.read_xe_file as

select object_name as event, CONVERT(xml, event_data) as data

from sys.fn_xe_file_target_read_file

('c:\wait_stats*.xel', 'c:\wait_stats*.xem', null, null)

go

create view dbo.xe_file_table as

select

      event

      , data.value('(/event/data/text)[1]','nvarchar(50)') as 'wait_type'

      , data.value('(/event/data/value)[3]','int') as 'duration'

      , data.value('(/event/data/value)[6]','int') as 'signal_duration'

from dbo.read_xe_file

go

One such case is doing the demo on multiple CPUs. Even though it appears as Resource Governor is "not working", it is "by design". Let's take a look at what happens in this case.

What is our promise about CPU bandwidth distribution?

When you specify percentage on the pool we will ensure average CPU bandwidth distribution on per scheduler basis for CPU bound workloads. Note, emphasis: average, per scheduler, CPU bound. What does this mean?

For example, you have a 2 CPU machine, and SQL Server runs on both of the CPUs. If you haven't set affinity mask as was indicated in the Setup step of the previous post,

will get you something like:

Important item to note here is that we have 2 online schedulers to service normal user requests (scheduler_id < 255). You can compare it with the case when affinity mask is set to 1. You will still have 2 schedulers (with scheduler_id < 255), but one of them will be offline.

Now, for simplicity, let's configure 2 groups and 2 pools as follows:

And corresponding classifier function as

Create logins, assign function and make changes effective 

Add group counters for CPU Usage % and run 1 CPU intensive query UserSlow, then 1 with UserFast. You will observe a picture similar to this:

But how can explain that the resource governor does not obey 70:30 distribution we specified for the resource pools? The answer is quite simple: the queries landed on different schedulers (= different CPUs) and thus, there is nothing to share between them and each of them is taking a full CPU. How can you verify that?

First of all, verify the classification 

Where you will see something like

Everything is normal here. Now, let's take a look at how currently executing requests landed on schedulers:

Which will get

What you can see here is that gFast and gSlow is working on different schedulers, thus confirming our earlier statement. Thus Red and Blue line corresponds to utilization of 1 full CPU or 50%, since it is normalized to number of CPUs available on the box.

Is this a problem with Resource Governor in general? Certainly not if you have more executing requests than you have CPUs or if queries are running in parallel. But you should keep this in mind if you are trying to give 90% of CPU bandwidth to a single query running serially compared to a parallel query or number of serial queries in another group/pool.

A simple way of thinking about this is: is my workload capable of taking as much CPU as I told it to? Remember, when you set MAX_CPU_PERCENT on the pool, you allowing workloads belonging to the pool to take this much bandwidth on each scheduler. In case of the demo, you are working with a single query workload which produces this initially unexpected result.  

Without further delay, let's get to point. As we have recently released a public CTP (Community Technology Preview) of SQL Server 2008 (aka SQL Server 2008 November CTP), which has Resource Governor functionality, and a number of presentations with Resource Governor demo on various conferences, I started getting a lot of questions about the demo itself which made me write this post.

We created this demo when Resource Governor was only a prototype to illustrate the concept, but probably its simplicity made it so attractive that it was included in a number of talks and demos by different people all over the world.

Background

If you are looking for general concepts about Resource Governor, check out the Books Online which comes as a separate download or read it directly on MSDN Library.

The Demo in a Nutshell

In the demo we will create two workload groups which share a common resource pool and one workload group which has its own dedicated resource pool. Each group gets a CPU intensive workload. Using the Resource Governor you will be able to affect CPU distribution between these competing workloads. The effects of the resource management are observed using performance counters specific to Resource Governor.

Step 1: Initial Demo Setup

On a newly installed server (i.e. no prior Resource Governor configuration) you may need to run these setup steps. I normally doing my demo on a dual core laptop and for the sake of simplicity I'm using a single CPU for SQL Server. To do so, I adjust CPU affinity mask as follows:

Using 1 CPU for SQL Server on a dual proc machine has an interesting side-effect: we normalize "CPU usage %" counter to number of CPUs on the box and thus, the values will hover around 50% as maximum and not 100% as you might expect. I will illustrate this below.

Of course you can use both CPUs but this complicates the demo. This is a subject for whole another post - what happens when you have multiple CPUs.

In addition to that for demo purposes I will set min/max server memory to a fixed value, since it will improve predictability of the demo on the laptop.

Step 2: Workload groups and Resource Pools

Now we will be setting up the following hierarchy of workload groups and resource pools:

Each corresponding workload group contains queries of the corresponding class or department (i.e. Marketing, Adhoc, and VP). Note that, Marketing and Adhoc queries share the same resource pool, while workload group VP has its own similarly named pool. The reason of such separation will become clear as we see how we adjust Resource Governor controls.

You will start building the above configuration in a bottom up manner (i.e. starting from pools and going up)

To do so, we execute the following T-SQL:

Step 3: Classification

Now, what you have just done is created hierarchy of the groups and pools, however, how does the server know about which query goes where? This is where classification comes in. The above picture becomes:

There is a couple of things:

1. To do the classification you will need to create a user-defined function that will be executed for every new connection and it will place these new connections in the corresponding workload groups.
2. How will we separate different connections? For demo purposes we will use 3 separate login names which we will check and use inside of the function

To implement the above 2 steps we will run the following:

Step 4: Are we there yet?

After all this work, can we start workloads and see what happens? The answer, as you have guessed by the question is - no. What's left? Again, a couple of steps:

1. We need to tell Resource Governor to use the function that we just created
2. Make all the changes effective

First step is done by

For the second step, let's compare output of catalog views with in-memory information (note difference in names of catalog views and dynamic management views (DMVs) which are prefixed with dm_:

Now transfer changes from metadata to memory by running the following statement. Also, do not confuse it with already existing RECONFIGURE command:

And rerun the above query on metadata and DMVs and you should see that new groups, pools and classifier function ID are present in corresponding DMVs.

Step 5: Running the workloads

The easiest way to simulate a CPU intensive workload is to run the following in a loop:

Also, instead of running this query from the Management Studio, consider saving it in a file and running from a command prompt by using a script similar to the below. Note that we are using 3 different user names to connect to the server.

To observe the effects of the load, add the following performance counters in the perfmon:

1. We will monitor CPU usage per group in the 1^st instance of perfmon; add "SQLServer:Workload Group Stats object", "CPU usage %" counter for "GroupMarketing", "GroupAdhoc" and "GroupVP" instances
2. We will monitor CPU usage per pool 2^nd instance of perfmon, add "SQLServer:Resource Pool Stats object", "CPU usage %" counter for "PoolMarketingAdhoc" and "GroupVP" instances

Before you start the next workload, observe the counters for pools and groups for a number of seconds, you should see approximately the following:

For groups (click on the image to open in a new window):

For pools:

Few things to note:

1. Remember I set affinity mask to 1? This is why the maximum on the figures is only 50% (it is for a single CPU while usage is normalized to all CPUs on the machine, 2 in my case)
2. At point A, I started Marketing workload, CPU usage went to the maximum on the CPU for a single workload, 50% in this case
3. At point B, I started VP workload, which, as you remember, belongs to a separate resource pool and it made Marketing group share half of the CPU with VP group. The same is true for the pools (because we have 1 to 1 match of active requests in groups to pools at this point)
4. At point C, I started Adhoc workload, which has its own group but shares the pool with Marketing workload. What happened here on the groups is that all 3 of them are now sharing the CPU getting approximately 1/3 of it or roughly 17%. On the pool side, however, we can see that MarketingAdhoc pool which shows aggregate resource usage by all groups inside of it, has 2/3 or the CPU while PoolVP has only 1/3. This behavior is exactly what you get on SQL Server 2005. Resources are distributed as they are being requested and whoever needs more CPU simply gets it. This was one of the goals: to make behavior as close as possible to SQL Server 2005 when you do not use the Resource Governor or adjust any parameters of it. Note, however, by creating the groups and pools you are already slightly altering the behavior of SQL Server 2008, but more on this in the next few posts.

Step 6: Management Actions

Now we came to the point where we want to apply action to change the above picture. Specifically, we want our VP workload to proceed faster and thus, limit CPU usage by Marketing and Adhoc workloads to 50% of the CPU.

To do this, we alter the PoolMarketingAdhoc using the following syntax (remember, we created the pool using all default parameters):

Remember to make changes effective:

Now, let's look at the counters:

For groups:

For pools:

What you will see is what happened at point D: Pool usage by PoolMarketingAdhoc went back to half of the CPU while PoolVP took the remaining part of it. Note that, on group side GroupVP usage went up to use half of the CPU (it has a single workload) while groups Marketing and Adhoc divide another half equally (12.5% each).

Further step is to alter IMPORTANCE parameters of the Marketing and Adhoc groups. IMPORTANCE affects CPU distribution when two groups share the same pool. Value of importance has a numeric meaning for CPU bandwidth distribution. Ratio of Low:Medium:High IMPORTANCE is equal to 1:3:9.

To illustrate this, let's run the following:

When changes become effective corresponds to the point E on the above graphs. Note that distribution between pools does not change, amount of CPU available to GroupVP also does not change (there are no other groups in this pool), but GroupMarketing takes about 10 times of the bandwidth available to GroupAdhoc (remember the ratio of Low:High = 1:9, but together they stay within PoolMarketingAdhoc limits.

That's enough for the first post. In the upcoming post I will try to go over the cases when the demo does not seem to work or shows "unusual" behavior and possible explanations of it.