In one of my first jobs as a full-fledged developer, I took over a project for a program that was suffering scaling issues. It would work reasonably well on small data sets, but would completely crash when given large quantities of data.
As I dug in, I found that the original programmer sought to speed things up by parallelizing the analysis - launching a new thread for each additional data source. However, he'd made a mistake in that all threads required a shared resource, on which they were deadlocking. Of course, all benefits of concurrency disappeared. Moreover it crashed most systems to launch 100+ threads only to have all but one of them lock. My beefy dev machine was an exception in that it churned through a 150-source dataset in around 6 hours.
So to fix it, I removed the multi-threading components and cleaned up the I/O. With no other changes, execution time on the 150-source dataset dropped below 10 minutes on my machine, and from infinity to under half an hour on the average company machine.
