Currently, serialized changes to all pages use the same mutex; we should try splitting that mutex up into multiple mutexes. The mutex logging output for one of a serialization function makes it look like the serialization mutex spends a fair amount of time blocking itself:

Because it's a pthread mutex which can be relatively large, we don't want to put one on every page, but we could create an array of pthread mutexes and hash, or round-robin assign, pages to mutexes.

@agorrod: what wtperf job mix would you suggest using for measuring this, what's the test that does updates to different pages?