I thought about this over the summer when I was working on 4027. Pretty much same idea as Nicholas here. I figured LIRs might be troublesome to implement - and also thought that newer features, such as 4027 or the reference counting patch, was a better use of time.
The larger the cache gets, the less the eviction algorithm matters. And we seem to be trending towards the possibility of larger caches. ~Li On Tue, Feb 21, 2012 at 9:01 AM, Nicolas Spiegelberg <[email protected]>wrote: > We had the author of LIRS come to Facebook last year to talk about his > algorithm and general benefits. At the time, we were looking at > increasing block cache efficiency. The general consensus was that it > wasn't an exponential perf gain, so we could get bigger wins from > cache-on-write intelligence, in-memory data compression techniques, and > adding stats so we could understand how to tune the existing LRU > algorithm. I still think that these 3 goals are more important at the > moment because LIRS would be a decent bit of code and only incremental > gain. It's probably something to revisit in a year or two. > > Nicolas > > On 2/21/12 8:26 AM, "[email protected]" <[email protected]> wrote: > > >Hi, > >Shall we experiment with low inter-reference recency set replacement > >policy to see if block cache becomes more effective ? > > > >Cheers > > > > > >
