Monday, May 9, 2011

Filing for a Cache Refund

My original plan was to publish this article on April 18th in celebration of Tax Filing Day.  However, life has a way of happening, so I filed IRSQL Form 1433 and applied for an extension.

Say Brother, can you spare me some Cache?

Have you used up all of your Plan Cache?  Would you like a refund?  Are ad-hoc queries bloating your Plan Cache?  How can you tell?  What are ad-hoc queries?  What’s a Plan Cache?

Ad-hoc Queries

Ad-hoc queries are TSQL queries that are issued directly against the database.  Typically, the query text is created directly in the application, then sent directly to the database for the data.  This is very common in web-based applications such as ASP or PHP.  However, you also see this with traditional, desktop application written in C/C++ or Java.

Contrast this with Stored Procedures.  Similar to procedures in other programming languages, Stored Procedures can be thought of as functions, methods, or procedures that are internal to SQL Server.  They are written in TSQL but stored within SQL Server.  The first time a Stored Procedure is executed, SQL Server parses and compiles it, and stores the resultant Execution Plan in the Plan Cache.  The Execution Plan tells SQL Server how it is going to get the data to satisfy the request.

The reason you want to cache Query Plans is that if you run a query once, chances are you will run it again.  Some queries may run thousands of times during the course of a day.  It all depends upon the application.

Plan Cache

The Plan Cache is the area of memory where SQL Server stores the Execution Plans for any queries or stored procedures.  Compiling and caching the execution plan does take a few extra CPU cycles away from the server, so why would you do it?  The next time that query or stored procedure is run, SQL Server already knows how to get the data in an efficient manner.  As your application scales, and your query or stored procedure is being run hundreds or thousands of time, this helps the execution to be a fast as possible.

Single-Use Plans

The goal with caching Query Plans is for the execution count be as high as possible.  This means we are getting the most bang for our buck.  However, if you have lots of single-use plans, then you are wasting CPU cycles calculating the plans, and you are wasting memory by storing them.

Every CPU cycle that is wasted compiling and caching a single-use Query Plan is a CPU cycle that could have been doing some work for your application.  Every byte of memory that is wasted storing a single-use Query Plan is memory that could have been used somewhere else.

If you have a high number of single-use plans in your cache, this could be contributing to high CPU utilization, excessive IO, or low memory on your server.

How to Find Ad-hoc Queries and Single-Use Plans

One way to find out if you have ad-hoc queries hitting your database is to ask the development teams if they have any ad-hoc queries in their code.  There are a couple of problems with this approach.

For one, they may not be aware if they are using ad-hoc queries.  If you have a large application, each developer may only familiar with a small part of it.  Also, if you have inherited an existing application, it may be difficult or impossible to wrap your head around the entire code base.

Another scenario is that the developers may be aware of some of their ad-hoc queries, but may underestimate the number of times they are being called, or they may not realize the impact they have upon the system.

As better approach is to ask SQL Server to tell you what sort of plans it has cached.

sys.dm_exec_cached_plans

SQL Server provides us with a Dynamic Management View, sys.dm_exec_cached_plans which lists out every Execution Plan that is currently stored in the Query Cache.  At the most basic level, you can query this DMV to see the details on every Query Plan in your system.

[sql]

-- basic query
select * from sys.dm_exec_cached_plans

[/sql]

By itself, this is not very useful. Of course, like all SQL queries you can add various WHERE clauses or aggregates in order to make it more meaningful for you.

So, how much are you wasting?

The way I approached this problem was to calculate how much memory the entire Plan Cache was using, then compare that to how much the Single-Use Adhoc Plans were using.  Once I had those numbers, I did some quick math to figure out what percentage of my Plan Cache was going to Single-Use Adhoc Plans.

[sql]

-- check percentage of the plan cache that is wasted on single-use adhoc plans

-- use a CTE to check across multiple servers
-- horizontal formatting
-- all plans
;with all_plans as (
select @@SERVERNAME as 'ServerName',
COUNT(*) as 'All Plans Count',
sum(cast(size_in_bytes as bigint)) / 1024 / 1024 as 'Total MB',
avg(cast(size_in_bytes as bigint)) / 1024 as 'Average KB'
from sys.dm_exec_cached_plans
),

-- adhoc plans
adhoc_plans as (
select @@SERVERNAME as 'ServerName',
COUNT(*) as 'Adhoc Plans Count',
sum(cast(size_in_bytes as bigint)) / 1024 / 1024 as 'Total MB',
avg(cast(size_in_bytes as bigint)) / 1024 as 'Average Bytes'
from sys.dm_exec_cached_plans
where objtype = 'Adhoc'
),

-- single-use adhoc plans
single_adhoc as (
select @@SERVERNAME as 'ServerName',
COUNT(*) as 'Single-Use Adhoc Plans Count',
sum(cast(size_in_bytes as bigint)) / 1024 / 1024 as 'Total MB',
avg(cast(size_in_bytes as bigint)) / 1024 as 'Average Bytes'
from sys.dm_exec_cached_plans
where usecounts = 1
and objtype = 'Adhoc'
)

-- query the CTEs

-- geek version, show all the data
select ap.[All Plans Count], ap.[Total MB], ap.[Average KB],
ah.[Adhoc Plans Count], ah.[Total MB], ah.[Average Bytes],
sa.[Single-Use Adhoc Plans Count], sa.[Total MB], sa.[Average Bytes],
cast(cast(sa.[Total MB] as numeric) / cast(ap.[Total MB] as numeric) * 100 as numeric(5,2)) as 'Percent Wasted on Single-Use Adhoc Plans'
from all_plans ap
join adhoc_plans ah
on ap.ServerName = ah.ServerName
join single_adhoc sa
on ah.ServerName = sa.ServerName

[/sql]

I also have an Executive Version, which shows the Percentage Wasted without adding in all those confusing details.  Replace the Geek Version with this one when you query the CTEs.

[sql]

-- executive summary, show how much memory is being wasted on single-use adhoc plans
select ap.[Total MB] as 'All Plans MB',
sa.[Total MB] as 'Single-Use Adhoc MB',
cast(cast(sa.[Total MB] as numeric) / cast(ap.[Total MB] as numeric) * 100 as numeric(5,2)) as 'Percent Wasted on Single-Use Adhoc Plans'
from all_plans ap
join adhoc_plans ah
on ap.ServerName = ah.ServerName
join single_adhoc sa
on ah.ServerName = sa.ServerName

[/sql]

Getting Buy In

If you encounter resistance to changing this setting, run this query against multiple servers and save the results in Excel.  Bosses love Excel; it helps them think.  :-)

Cache Refund

Even better, write up a simple Word Document or PowerPoint Presentation and tie it back to the business or the budget.  That will make an impression.

How much did you spend on memory when you built the server?  Do you see evidence of memory pressure on your server?  Are some of your key business processes running slow because they are not able to get enough memory.

One technique I have used is to put your conclusion first.  Then, lead them along the path that led you to your conclusion.  Put in a few key graphs and screenshots, and voila!  Many managers will not read beyond the first slide or paragraph so you need to work on getting your core message conveyed in a very brief space.

How to fix it

The fix is quite simple: optimize for adhoc workloads.

[sql]

-- show advanced options - enables all options to be seen / set
use [master]
go
exec sp_configure 'show advanced options', 1
go
reconfigure
go

-- check to see if it is turned on
sp_configure 'optimize for ad hoc workloads'
go

-- optimize for ad hoc workloads
sp_configure 'optimize for ad hoc workloads', 1
go
reconfigure
go

-- verify that it has been set correctly
sp_configure 'optimize for ad hoc workloads'
go

[/sql]

This is an Instance Level setting that tells SQL Server not to compile and cache the Query Plan the first time it encounters an Adhoc Query.  Instead, it compiles a Plan Stub.

The Plan Stub is a small identifier that allows SQL Server to recognize that a particular plan has been looked at before.  This way, if it does end up getting multiple uses down the road, SQL Server is able to compile and cache the plan as normal.

For many of my servers, the Plan Stub was an Order of Magnitude smaller than the size of the full Query Plan; 320 bytes versus 25,000k-35,000 bytes for the average Single-Use Adhoc Query Plan.  Of course, your mileage may vary.

Make it so, Number One!

Something to keep in mind, is that you will not be able to predict or control what SQL Server will decide to use its newfound memory on.  Kimberly Tripp (blog | twitter) has an excellent article which discusses some of the behaviors you may observe.

Conclusion

Taking a look inside your Plan Cache can be an illuminating experience.  While you may not get any money back, it is worth looking into.

1 comment:

  1. [...] You may see my CTE version of this code by viewing yesterday’s blog post. [...]

    ReplyDelete