In a simple, nonassuming patch for Oracle8i Release 8.1.5, Oracle introduced a completely altered LRU chain algorithm that has virtually eliminated all LRU chain latch contention problems. Amazingly, it works without tweaking instance parameters, was slid into the world's largest production Oracle systems without any announcement, and was not documented anywhere in official Oracle documentation. Even scouring Oracle Database 11g documentation, I could find only references to the LRU chain algorithm as a modified LRU.

I first discovered this algorithmic change while looking at a list of instance parameters and noticing some new strange ones, like _db_percent_hot_default and _db_aging_cool_count. I had never seen these parameters before, and as I have learned, whenever there are new parameters or old parameters suddenly disappear, an algorithm has been changed. At that time, I had no idea just how significant this change was. I started checking my typical sources4 and discovered that Oracle had indeed implemented what is commonly called in the computer science field the count-frequency scheme.

As you might expect, the general approach is to increment a counter every time a buffer header is touched. The more popular buffer headers will have a higher touch count and are deemed more popular, and therefore should remain in the buffer cache. While this is generally true, how it is implemented is quite different. Figure 6-18 shows a diagram of the touch-count algorithm.