Optimization has become one of my favorite topics. OK, I’ll admit it: it might be an obsession and a source of joy. My inability to optimize (queries, data models, and DAX) used to be a source of headaches and the occasional heart attack depending on what I was trying to accomplish. I make no claim that I’m an optimization expert, but I’ve spent hours wondering and researching why a data model’s calculation time could go from a couple of seconds to many minutes. This is a rare occurrence, but when it has happened, it’s happened when I’ve been on the verge of something great. Of course, if you’re taking business intelligence to its edge, you’re not adding 2+2—you’re trying to give someone an answer they (and you) thought was impossible. A recalculation that takes many minutes (especially if you’re not the end-user) is unacceptable and may cause your data model to crash—fatal exceptions anyone?
This article focuses on an optimization technique that I couldn’t find anywhere. Before I get into that technique, I’d like to touch on what I consider to be the basic tenets of optimization when working with data modeling and DAX.
Read on for the technique.
Now that the analysis database contains the performance data, you can use WorkloadViewer to visualize it and draw your conclusions.
WorkloadViewer is a GUI tool that reads performance data from the analysis database and gives a graphical representation using charts and grids. It accepts a number of command line arguments that allow to automate its behavior, but it can be also opened without specifying any arguments: in this case, WorkloadViewer will present a form to fill the missing information.
WorkloadViewer can be used to visualize information about a single benchmark (analysis mode) or two benchmarks (comparison mode). In this case, you just need to work with a single benchmark, so it is enough to enter the connection info on the left, including the schema name where the tables are. When using Windows Authentication, you can leave UserName and Password blank.
Gianluca has a full demo from the beginning of data capture to analysis.
It turns out that on one of the benchmarks, Microsoft was able to achieve higher throughput by setting the SQL Server process to
HIGH_PRIORITY_CLASS, and the thread priority level over and above that to
THREAD_PRIORITY_HIGHEST. While it was extremely helpful to beat artificial benchmarks, it came at the cost of giving SQL Server higher execution context on the CPU than almost every other process on Windows.
And that’s the only good use of priority boost ever. If you’ve never heard of priority boost, just keep on ignoring it.
This blog post will examine changes to the query plan and performance when Scalar UDF Inlining is occurring.
I have previously blogged about function performance – here and here. For a quick recap, the performance test ranks these function in duration. The order of the types of functions by duration is Inline TVF, Scalar UDF, and then finally a Multi-Statement TVF (MSTVF) – and the MSTVF is way behind the other two types of functions.
I’m using a Linux (Ubuntu) VM with SQL Server 2019 to perform these comparison performance tests. I use one database in the SQL 2019 compatibility level, and another one in the SQL 2017 compatibility level. I’m using the same performance test used in the previous blog posts.
Wayne finds a definite performance improvement, but not enough in my mind to start creating a bunch of these.
At first, this statement might sound a bit confusing. Usually, we expect wait statistics to show us what a query is waiting on, whether it’s waiting on memory, loading pages from disk, or any of the other numerous wait types.
Once you start collecting wait statistics, you’ll have a lot of data to sort through. You might find waits like CX_PACKET, CX_CONSUMER, and PAGEIOLATCH. Surprisingly, these could mean that your databases aren’t well indexed.
This makes sense. At its core, wait stats tell you where SQL Server is hurting: where is the bottleneck. But just like a person at the doctor, SQL Server can only be so specific in how it relates this pain to you, and that specificity generally boils down to hardware components. The solution might be “get more hardware,” but as Arthur points out, writing better queries and using better indexes can mitigate those pains too.
Yes, Entity Framework will improve your job quality and reduce stress in your life.
With one caveat, it gets used correctly.
That’s the hard part right? There is tons of technology that makes things better, if used correctly. There are all sorts of programs that make your life easier, if used correctly. Yet, all of these, used incorrectly, can make your life a hell.
One nit that I’ve always had with Entity Framework is that it’s very difficult to tell what part of the code the call was coming from. You really have no idea. So when my friend, Chris Woodruff, asked me on Twitter what would be the best way to monitor TagWith queries in Entity Framework, well, first, I had to go look up what TagWith was, then I got real excited, because, hey, here’s a solution.
That “I love Entity Framework” is the lead-in to a one-act play of mine with people with pitchforks, tar, and feathers. Nevertheless, Grant shows us how we can tag code in C# and capture that data in extended events. I’d read it but I’m too busy sharpening my pitchfork.
Since, in spite of our best efforts since SQL Server 2000, we can’t effectively stop scalar UDFs from being used, wouldn’t it be great to make SQL Server simply handle them better?
SQL Server 2019 introduces a new feature called Scalar UDF Inlining. Instead of keeping the function separate, it is incorporated into the overall plan. This leads to a much better execution plan and, in turn, better runtime performance.
Read the whole thing. Especially the part about how this changes the way one of the DMV counters behaves.
So according to this DMV, there’s only 3 queries with memory grants, while the remaining 5 queries have to wait for space in this semaphore. This is where the wait type comes in. When a query is sitting as a waiter, it will display the wait type RESOURCE_SEMAPHORE.
This is one of the biggest performance-killing waits you can find, and there are several ways to tackle it in SQL Server, as well as adding more hardware.
Last month, I covered a puzzle involving matching each row from one table with the closest match from another table. I got this puzzle from Karen Ly, a Jr. Fixed Income Analyst at RBC. I covered two main relational solutions that combined the APPLY operator with TOP-based subqueries. Solution 1 always had quadratic scaling. Solution 2 did quite well when provided with good supporting indexes, but without those indexes also had quadric scaling. In this article I cover iterative solutions, which despite being generally frowned upon by SQL pros, do provide much better scaling in our case even without optimal indexing.
Itzik has three separate solutions here, including one using the CLR.
The gist of this kit is that it is a database repository as part of the sp_BlitzFirst to collect monitoring alerting and performance metric data. Once you’ve set this up, then you can use a Power BI desktop dashboard as an interface for all that data.
Now this is an awesome way to introduce more DBAs to Power BI and it’s a great way to get more out of your metrics data. The challenge is, it’s a lot of data to be performing complete refreshes on and the natural life of a database like this is growth. The refresh on the static database I was sent by Tracy, once I connected my PBIX to the local db sources, took upwards of an hour to refresh. Keep in mind, I have 16G of memory, 32G of swap and have upped my data load options in Power BI quite high.
Kellyn walks through the things she did to improve performance as a starting point, so check it out and be aware that there’s even more that can be done.