What is Parallelism and why should I care?

You know that wait stat you sometimes see; ‘CXPACKET’? You know all of that blocking you’re getting on your server? You know of those times when a user is waiting on the system and there’s nothing obvious causing it? That’s why you should care. Now, parallelism isn’t a silver bullet to fix all of your problems but the standard settings for this are so ancient that this could be causing you major headaches.

 

First of all, a quick pop quiz

In SSMS, right click on your server and go to properties. Open the ‘Advanced’ tab and scroll to the bottom. Take a look at Cost Threshold for Parallelism and Max Degree of Parallelism. I bet this is set to 5 and 0 right? If not, good for you, but do you understand these settings? If you are seeing a 5 and a 0, don’t worry, these are the defaults but you should be aware that they’re rubbish and you’ll want to be changing these.

 

Parallelism Basics

Big queries are hungry, they take a lot of CPU to process, it would be great if we could split up the work that they can do so they run faster right? Yep, it’s a great idea. Instead of using a single core of that quad core processor your server is running on, wouldn’t it be a great idea if we could use more and run our query in parallel? Yep, but there’s some gotchas. Let’s talk about these two settings before we think about changing them.

 

Cost Threshold for Parallelism

Let’s get the first question out of the way, what does this number stand for? What exactly is (the default) of 5? Is it seconds?

• Short answer: Yes but no.
• Long answer: it was a count of seconds a long time ago in a land far far away (Microsoft HQ, Washington). Now it’s just a number that we can use to compare the relative size of queries.

This number is calculated by the optimizer when a query is compiled. Statistics are compared and SQL decides how much it thinks the query is going to cost. If the threshold is over this number then it’s going to try to run parallel. But how many cores is it going to use?

 

Max Degree of Parallelism (MAXDOP)

Optimizer has decided that the query it’s going to run is big enough to run parallel, how parallel it goes is based upon this setting. The SQL default is set to 0 which for this setting means unlimited. This means that this massive query we have is going to use as many cores as it likes, on that quad core processor it’s going to split itself into 4 and use all of the CPU so that it can run as fast as it can. Sounds great right, this big query is going to run really fast? Not in the real world.

 

Why are the defaults rubbish?

Technology has come quite a way since SQL was designed. What used to be a large query is, comparatively, tiny by today’s standards. Queries with a cost of 5 or more is likely to be quite a number of queries, chances are that you’re going to have relatively small queries going parallel. When they do, your MAXDOP being set to 0 (unlimited) is going to mean that the majority of your queries are going to be taking up the whole CPU every time they are run. They’re going to be fighting over your resource which is going to be killing your performance and it likely to be a cause of all of those blocks that you’re seeing.

 

Damn, how do I fix it?

Great thing is, this can be a quick win but it might take a bit of tinkering to get the best possible settings for your application.

• A good place to start for the Cost Threshold setting is to stick it to 50 to begin with. You may feel that you want to reduce or increase this setting as time moves on but give it a go.
• MAXDOP will require a bit of working out for you, but only a bit! How many cores does your server have on the CPU? 4? 8? Try setting your MAXDOP to half of the number of available cores up to a maximum of 4 without hyper threading and 8 with hyper threading turned on.

Don’t leave these settings and walk away. They’ll be fine, but what if a little tweak helps further? Continue to monitor your blocking to see if this alleviates the problem.