In the TODO list is this item: *Modify the planner to better estimate caching effects * Tom mentioned this in his presentation at PGCON, and I also chatted with Tom about it briefly afterwards.
Based on last year's discussion of this TODO item, it seems thoughts have been focused on estimating how much data is being satisfied from PG's shared buffers. However, I think that's only part of the problem. Specifically, read performance is going to be affected by: 1. Reads fulfilled from shared buffers. 2. Reads fulfilled from system cache. 3. Reads fulfilled from disk controller cache. 4. Reads from physical media. #4 is further complicated by the type of physical media for that specific block. For example, reads that can be fulfilled from a SSD are going to be much faster than ones that access hard drives (or even slower types of media.) System load is going to impact all of these as well. Therefore, I suggest that an alternative to the above TODO may be to gather performance data without knowing (or more importantly without needing to know) which of the above sources fulfilled the read. This data would probably need to be kept separately for each table or index, as some tables or indexes may be mostly or fully in cache or on faster physical media than others, although in the absence of other data about a specific table or index, data about other relations in the same tablespace might be of some use. Tom mentioned that the cost of doing multiple system time-of-day calls for each block read might be prohibitive, it may also be that the data may also be too coarse on some systems to be truly useful (eg, the epoch time in seconds.) If this data were available, that could mean that successive plans for the same query could have significantly different plans (and thus actual performance), based on what has happened recently, so these statistics would have to be relatively short term and updated frequently, but without becoming computational bottlenecks. The problem is one I'm interested in working on. -- Mike Nolan
