A few years ago, I had a problem.  A database routine for processing bulk payments for LedgerSMB that was humming along in my test cases was failing when tested under load with real data prior to implementation.  Some testing showed that while it ran fine with small numbers of inputs, it eventually slowed way down as the number of inputs rose.  Andrew Sullivan, who was also on the project through a different company, suggested it was something called a "cache miss" because I was running through a loop, inserting two rows, and then updating a different table.    Al...