SQL Server 2012 : SQL Server Architecture - SQL SERVER’S EXECUTION MODEL AND THE SQLOS

So far, this article has abstracted the concept of the SQLOS to make the flow of components through the architecture easier to understand without going off on too many tangents. However, the SQLOS is core to SQL Server’s architecture so you need to understand why it exists and what it does to complete your view of how SQL Server works.

In short, the SQLOS is a thin user-mode layer that sits between SQL Server and Windows. It is used for low-level operations such as scheduling, I/O completion, memory management, and resource management. To explore exactly what this means and why it’s needed, you first need to understand SQL Server’s execution model.

Execution Model

When an application authenticates to SQL Server it establishes a connection in the context of a session, which is identified by a session_id (in older versions of SQL Server this was called a SPID). You can view a list of all authenticated sessions by querying the sys.dm_exec_sessions DMV.

When an execution request is made within a session, SQL Server divides the work into one or more tasks and then associates a worker thread to each task for its duration. Each thread can be in one of three states (that you need to care about):

• Running — A processor can only execute one thing at a time and the thread currently executing on a processor will have a state of running.
• Suspended — SQL Server has a co-operative scheduler (see below) so running threads will yield the processor and become suspended while they wait for a resource. This is what we call a wait in SQL Server.
• Runnable — When a thread has finished waiting, it becomes runnable which means that it’s ready to execute again. This is known as a signal wait.

If no worker threads are available and max worker threads has not been reached, then SQL Server will allocate a new worker thread. If the max worker threads count has been reached, then the task will wait with a wait type of THREADPOOL until a thread becomes available. Waits and wait types are covered later in this section.

The default max workers count is based on the CPU architecture and the number of logical processors. The formulas for this are as follows:

For a 32-bit operating system:

• Total available logical CPUs <= 4
  • Max Worker Threads = 256
• Total available logical CPUs > 4
  • Max Worker Threads = 256 + ((logical CPUs - 4)*8)

For a 64-bit operating system:

• Total available logical CPUs <= 4
  • Max Worker Threads = 512
• Total available logical CPUs > 4
  • Max Worker Threads = 512 + ((logical CPUs - 4)*16)

As an example, a 64-bit SQL Server with 16 processors would have a Max Worker Threads setting of 512 + ((16–4)*16) = 704.

You can also see the max workers count on a running system by executing the following:

SELECT max_workers_count
FROM sys.dm_os_sys_info

Each worker thread requires 2MB of RAM on a 64-bit server and 0.5MB on a 32-bit server, so SQL Server creates threads only as it needs them, rather than all at once.

The sys.dm_os_workers DMV contains one row for every worker thread, so you can see how many threads SQL Server currently has by executing the following:

SELECT count(*) FROM sys.dm_os_workers

Schedulers

Each thread has an associated scheduler, which has the function of scheduling time for each of its threads on a processor. The number of schedulers available to SQL Server equals the number of logical processors that SQL Server can use plus an extra one for the dedicated administrator connection (DAC).

You can view information about SQL Server’s schedulers by querying the sys.dm_os_schedulers DMV.

Figure 1 illustrates the relationship between sessions, tasks, threads, and schedulers.

FIGURE 1

Windows is a general-purpose OS and is not optimized for server-based applications, SQL Server in particular. Instead, the goal of the Windows development team is to ensure that all applications, written by a wide variety of developers inside and outside Microsoft, will work correctly and have good performance. Because Windows needs to work well in a broad range of scenarios, the development team is not going to do anything special that would only be used in less than 1% of applications.

For example, the scheduling in Windows is very basic to ensure that it’s suitable for a common cause. Optimizing the way that threads are chosen for execution is always going to be limited because of this broad performance goal; but if an application does its own scheduling then there is more intelligence about what to choose next, such as assigning some threads a higher priority or deciding that choosing one thread for execution will avoid other threads being blocked later.

The basic scheduler in Windows is known as a pre-emptive scheduler and it assigns slices of time known as quantums to each task to be executed. The advantage of this is that application developers don’t have to worry about scheduling when creating applications; the downside is that execution can be interrupted at any point as Windows balances execution requests from multiple processes.

All versions of SQL Server up to and including version 6.5 used the Windows scheduler to take advantage of the work that the Windows team had done through a long history of optimizing processor usage. There came a point, however, when SQL Server 6.5 could not scale any further and it was limited by the general-purpose optimizations of the pre-emptive scheduler in Windows.

For SQL Server 7.0, Microsoft decided that SQL Server should handle its own scheduling, and created the User Mode Scheduler (UMS) to do just that. The UMS was designed as a co-operative scheduling model whereby threads aren’t forcibly interrupted during execution but instead voluntarily yield the processor when they need to wait for another resource. When a thread yields the processor, a wait type is assigned to the task to help describe the wait and aid you in troubleshooting performance issues.

The SQLOS

Prior to SQLOS (which was first implemented in SQL Server 2005), low-level operations such as scheduling, I/O completion, memory management, and resource management were all handled by different teams, which resulted in a lot of duplication of effort as the product evolved.

The idea for SQLOS was to consolidate all these efforts of the different internal SQL Server development teams to provide performance improvements on Windows, putting them in a single place with a single team that can continue to optimize these low-level functions. This enables the other teams to concentrate on challenges more specific to their own domain within SQL Server.

Another benefit to having everything in one place is that you now get better visibility of what’s happening at that level than was possible prior to SQLOS. You can access all this information through dynamic management views (DMVs). Any DMV that starts with sys.dm_os_ provides an insight into the workings of SQLOS, such as the following:

• sys.dm_os_schedulers — Returns one row per scheduler (remember, there is one user scheduler per logical processor) and displays information about scheduler load and health.
  
• sys.dm_os_waiting_tasks — Returns one row for every executing task that is currently waiting for a resource, as well as the wait type
• sys.dm_os_memory_clerks — Memory clerks are used by SQL Server to allocate memory. Significant components within SQL Server have their own memory clerk. This DMV shows all the memory clerks and how much memory each one is using.
  

Relating SQLOS back to the architecture diagrams shown earlier, many of the components make calls to the SQLOS in order to fulfill low-level functions required to support their roles.

Just to be clear, the SQLOS doesn’t replace Windows. Ultimately, everything ends up using the documented Windows system services; SQL Server just uses them in a way optimized for its own specific scenarios.