CXPACKET has to be most popular one of all wait stats. I have commonly seen this wait stat as one of the top 5 wait stats in most of the systems with more than one CPU.

Books On-Line:

Occurs when trying to synchronize the query processor exchange iterator. You may consider lowering the degree of parallelism if contention on this wait type becomes a problem.

CXPACKET Explanation:

When a parallel operation is created for SQL Query, there are multiple threads for a single query. Each query deals with a different set of the data (or rows). Due to some reasons, one or more of the threads lag behind, creating the CXPACKET Wait Stat. There is an organizer/coordinator thread (thread 0), which takes waits for all the threads to complete and gathers result together to present on the client’s side. The organizer thread has to wait for the all the threads to finish before it can move ahead. The Wait by this organizer thread for slow threads to complete is called CXPACKET wait.

Note that not all the CXPACKET wait types are bad. You might experience a case when it totally makes sense. There might also be cases when this is unavoidable. If you remove this particular wait type for any query, then that query may run slower because the parallel operations are disabled for the query.

Reducing CXPACKET wait:

We cannot discuss about reducing the CXPACKET wait without talking about the server workload type.

OLTP: On Pure OLTP system, where the transactions are smaller and queries are not long but very quick usually, set the “Maximum Degree of Parallelism” to 1 (one). This way it makes sure that the query never goes for parallelism and does not incur more engine overhead.

EXEC sys.sp_configure N'cost threshold for parallelism', N'1'
GO
RECONFIGURE WITH OVERRIDE
GO

Data-warehousing / Reporting server: As queries will be running for long time, it is advised to set the “Maximum Degree of Parallelism” to 0 (zero). This way most of the queries will utilize the parallel processor, and long running queries get a boost in their performance due to multiple processors.

EXEC sys.sp_configure N'cost threshold for parallelism', N'0'
GO
RECONFIGURE WITH OVERRIDE
GO

Mixed System (OLTP & OLAP): Here is the challenge. The right balance has to be found. I have taken a very simple approach. I set the “Maximum Degree of Parallelism” to 2, which means the query still uses parallelism but only on 2 CPUs. However, I keep the “Cost Threshold for Parallelism” very high. This way, not all the queries will qualify for parallelism but only the query with higher cost will go for parallelism. I have found this to work best for a system that has OLTP queries and also where the reporting server is set up.

Here, I am setting ‘Cost Threshold for Parallelism’ to 25 values (which is just for illustration); you can choose any value, and you can find it out by experimenting with the system only. In the following script, I am setting the ‘Max Degree of Parallelism’ to 2, which indicates that the query that will have a higher cost (here, more than 25) will qualify for parallel query to run on 2 CPUs. This implies that regardless of the number of CPUs, the query will select any two CPUs to execute itself.

EXEC sys.sp_configure N'cost threshold for parallelism', N'25'
GO
EXEC sys.sp_configure N'max degree of parallelism', N'2'
GO
RECONFIGURE WITH OVERRIDE
GO


Read all the post in the Wait Types and Queue series. Additionally a must read comment of Jonathan Kehayias.

Note: The information presented here is from my experience and I no way claim it to be accurate. I suggest you all to read the online book for further clarification. All the discussion of Wait Stats over here is generic and it varies from system to system. It is recommended that you test this on the development server before implementing on the production server.

Reference: Pinal Dave (http://blog.SQLAuthority.com)