At Tue, 11 Nov 2008 19:48:04 -0500, Kevin Darcy wrote: > > Obviously I'm not making myself very clear. > > I'm talking about a box with plenty of memory available, where a lot of > the commonly-accessed data persists in cache, and the goal is to make > the turnaround time for queries of this cached data to be > _consistently_short_. A time-sensitive app might be negatively impacted > if named decides to kick off a cleaning operation while the app is > trying to resolve a bunch of names. > > One option is to turn off cache-cleaning altogether. But eventually, I > would think, the memory structures would accumulate junk to the point > that performance would be impacted anyway. There should be a "sweet > spot" -- enough cleaning to keep cache fetches efficient, but not so > much as to give inconsistent query turnaround times because of > cache-cleaning overhead. Being able to schedule cleaning for times of > the day/week/month that are known to be low-volume or low-impact (which > aren't necessarily the same thing), would be ideal.
See the LRU cache cleaner that shipped in 9.5.0. It was a substantial rewrite of the cache cleaning mechanism, intended to address essentially this problem, but in a different way than you suggested. > In any case, I just threw that out there as a "would be nice" idea. We > don't have any pressing requirement for this, and I'm not aware that > anyone else does either... Understood, thanks.