Find an updated version of this post here:
Parallelism and MAXDOP
There’s no doubt that parallelism in SQL is a great thing. It enables large queries to share the load across multiple processes and get the job done quicker.
However it’s important to understand that it has an overhead. There is extra effort involved in managing the separate streams of work and synchronising them back together to – for instance – present the results.
That can mean in some cases that adding more threads to a process doesn’t actually benefit us and in some cases it can slow down the overall execution.
We refer to the number of threads used in a query as the DOP (Degree of Parallelism) and in SQL Server we have the setting MAXDOP (Maximum Degree of Parallelism) which is the maximum DOP that will be used in executing a single query.
Microsoft generally recommend caution setting MAXDOP above 8:
Here’s a nice post here from Kendra Little talking about how having higher settings can actually slow down your query execution time:
Out of the box, MAXDOP is set to 0, which means there is no limit to the DOP for an individual query. It is almost always worth changing this to a more optimal setting for your workload.
Often in tuning a SQL Server instance we will look at wait stats – which tell us what queries have been waiting for when they run. CXPACKET waits are usually associated with parallelism and particularly the case where multi-threaded queries have been stuck waiting for one or more of the threads to complete – i.e. the threads are taking different lengths of time because the load hasn’t been split evenly. Brent Ozar talks about that here:
High CXPACKET waits can be – but aren’t necessarily – a problem. You can cure CXPACKET waits by simply setting MAXDOP to 1 at a server level (thus preventing parallelism) – but this isn’t necessarily the right solution. Though in some cases in can be, SharePoint for instance is best run with MAXDOP set to 1.
What you can definitely deduce from high CXPACKET waits however is that there is a lot of parallelism going on and that it is worth looking at your settings.
Cost Threshold for Parallelism
This is another setting available to us in SQL Server and defines the cost level at which SQL will consider a parallel execution for a query. Out of the box this is set to 5 which is actually a pretty low number. Query costing is based on Algorithm’s from “Nick’s machine” the box used by the original developer who benchmarked queries for Microsoft.
Compared to modern servers Nick’s machine was pretty slow and as the Cost Threshold hasn’t changed for many years, it’s now generally considered too low for modern workloads/hardware. In reality we don’t want all our tiny queries to go parallel as the benefit is negligible and can even be negative, so it’s worth upping this number. Advice varies but generally recommendations say to set this somewhere in the range from 30 to 50 (and then tuning up and down based on your production workload).
There are many articles in the SQL Server community about how the out of the box setting is too low, and asking Microsoft to change it. Here’s a recent one:
In tuning parallelism we need to think about how we want different sized queries to act on our server.
In general we don’t want these to go parallel so we up the Cost Threshold to an appropriate number to avoid this. As discussed above 30 is a good number to start with. You can also query your plan cache and look at the actual costs of queries that have been executed on your SQL Instance to get a more accurate idea of where you want to set this. Grant Fritchey has an example of how to do that here:
As he mentions in the post, this assessment can be quite expensive to run – so do it when things are quiet.
Medium to Large Queries
This is where we want to take advantage of parallelism, and do so by setting MAXDOP to an appropriate level. Follow the guidelines from the Microsoft article referenced above. Here it is again:
Often the answer is going to be simply to set it to 8 – but then experiment with tuning it up and down slightly to see whether that makes things better or worse.
Very Large Queries
If we have a mixed workload on our server which includes some very expensive queries – possibly for reporting purposes – then we may want to look at upping the MAXDOP for these queries to allow them to take advantage of more processors. One thing to consider though is – do we really want these queries running during the day when things are busy? Ideally they should run in quieter times. If they must run during the day, then do we want to avoid them taking over all the server power and blocking our production workload? In which case we might just let them run at the MAXDOP defined above.
If we decide we want to let them have the extra power then we can over-ride the server MAXDOP setting with a query hint OPTION(MAXDOP n):
You will want to experiment to find the “best” value for the given query. As discussed above and as shown in Kendra Little’s article, just setting it to the maximum number of cores available isn’t necessarily going to be the fastest option.
Parallelism is a powerful tool at our disposal, but like all tools it should be used wisely and not thrown at every query to its maximum – and this is often what happens with the out of the box settings on SQL Server. Tuning parallelism is not a knee-jerk reaction to high CXPACKET waits, but something we should be considering carefully in all our SQL Server implementations.
I personally have seen numerous instances where a server is going crazy, timing out all over the place, where changing the parallelism settings has instantly brought everything back to earth, operating smoothly and quickly once more.