The Art of Performance Tuning

Performance tuning often gets called an art as people feel that a certain knack or innate talent comes into play. And while I don’t disagree that a certain level of knowledge is involved, I completely disagree that only certain people can performance tune.

Given the correct approach, anyone should be able to learn to effectively performance tune.

But how should performance tuning be approached? I’d like to take a step back from delving into specifics and define the process of performance tuning, a set of guidelines that can be used in any circumstance.

Performance tuning should follow the scientific method, which is defined as: “A set of principles and procedures for the systematic pursuit of knowledge involving the recognition and formulation of a problem, the collection of data through observation and experimentation and the formulation and testing of hypotheses.”

In practice, this can be broken down into the following steps:

•Observation

•Declaration of the end goal or issue

•Forming a hypothesis

•Testing the hypothesis

•Analysis of results

•Conclusion

•Further research

This way of approaching performance tuning comes into its own, particularly when investigation is required to respond to issues in a live environment. The trick is to follow the method and not to go with “gut” feelings or guesses. Stick to what you know.

For example, your production server has started responding slowly. Queries are taking longer than expected to complete. This could be due to memory pressure, but you are not sure. It would be better to start off by declaring “Performance of the server is poor,” which you know to be 100 percent correct, rather than declaring “Performance of the server is poor due to memory pressure.”

So the steps to follow would be:

•Observation: Queries on production running for longer than expected

•Declaration: Performance of the server is poor

•Hypothesis: The poor performance of the server is due to memory pressure

•Testing: Running an extended events session on the server to catch incoming queries

•Analysis: Several queries performing scans of large tables

•Conclusion: Creation of a covering non-clustered index to prevent the table scans

•Further Research: Are there any other queries being executed that are performing table scans?

These steps may seem obvious, but using the scientific method will prevent you from being led down blind alleys and potentially missing the actual problem. It provides a structure for the investigation.