How to eliminate impact from system itself on a performance test?

I am not quite sure whether this is a good place to ask this question, so please feel free to close it if an admin thinks it's inappropriate.

Basically, I am doing database update. We have a few builds per day for database development. Before we use the builds to update our prod databases, we do a performance test on some test databases to see how long it would take to update a database from one build to another.

All the test databases (5 in total, each is close to 20G) were restored from the same bak file. But we've observed that update duration on those test databases vary a lot from seconds to tens of seconds (we thought they should be closer considered all those test databases are from the same bak). So we suspect the system(or you can call it database server) might also ran some backend jobs while updating which delayed the update. We would like to find a way to eliminate the system impact or at least find out a statistical/mathematical solution to offset the delay made from the system.