[jbosscache-dev] eviction policies

[Manik Surtani]

On 8 Feb 2007, at 19:19, Mircea Markus wrote:

> Hi guys,
>
> I've walked around eviction policies and here are some thoughts
>
> 1) There is an eviction policy algorithm more efficient than LRU  
> (which I found quite popular) - Adaptive Replacement Policy. The  
> basic idea is to not rely only on the time of last access to the  
> node, but also on the number of time(frequency) a given node was  
> accessed. Here it is a nice description of how it works: http:// 
> en.wikipedia.org/wiki/Adaptive_Replacement_Cache.

This actually is quite interesting.  Something to look into.



>
> 2) One other type of eviction which I found handy is a time based  
> eviction, i.e. the node will be removed from the cache after a  
> certain time(countdown starts when it is added).  A common use case  
> for this is when one needs an info from database, which might be  
> changed, but is not critical to have it visible in real time.

I believe we do time-based eviction - FIFO with maxAge.

>
> 3) Memory based eviction policy. Very useful in the case in which  
> cache is used for storing very heterogeneous objects - from the  
> size point of view. If the objects are large one is assured that  
> the cache won't exhaust system memory and in the case of small  
> objects the cache manages to store much more than it does if it  
> would use a static defined policy, like LRU. The most appropriate  
> way (I found) for measuring object size is shipped within JDK5,  
> through Instrumentation.getObjectSize. The MAJOR drawback is  
> performance: method returns only the size of the base object; for  
> all aggregated objects one will have to iterate recursively over  
> the object graph and calculate size - costly both as performance  
> and as memory. I don't have the complete picture on the JBC usage  
> scenarios ( e.g. this might be useful in a scenario in which  
> addition is rare and memory size is critical) so perhaps you guys  
> can point out whether or not  this is feasible. Perhaps in the case  
> of PojoCache in which object are introspected anyway?

Like you said getObjectSize is expensive... so not very good if you  
have a lot of short-lived objects in mem.  OTOH, like you said for  
PojoCache where objects are broken down, it may not be that  
expensive.  Sizes could be calculated once when objects are put into  
the cache, for example.

>
> Any thoughts?
>
> Cheers,
> Mircea
>
>
> _______________________________________________
> jbosscache-dev mailing list
> jbosscache-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/jbosscache-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/jbosscache-dev/attachments/20070209/c7704245/attachment.html