javahashmapweak-referencessoft-references

Is there a SoftHashMap in Java?

I know there is a WeakHashMap in java.util, but since it uses WeakReferences for everything, which is only referenced by this Map, referenced objects will get lost on the next GC cycle. So it's nearly useless if you want to cache random data, which is very likely to be requested again without being Hard-linked the rest of the time. The best solution would be a map, which uses SoftReferences instead, but I didn't find one in the Java RT Package.

Solution

  • Edit (Aug. 2012):

    It turns out that currently the best solution are probably Guava 13.0's Cache classes, explained on Guava's Wiki - that's what I'm going to use. It even supports building a SoftHashMap (see CacheBuilder.newBuilder().softKeys()), but it is probably not what you want, as Java expert Jeremy Manson explains (below you'll find the link).

    Not that I know of (Nov. 2008), but you kind find some implementation of SoftHashMap on the net.

    Like this one: SoftHashMap or this one.

    Edit (Nov. 2009)
    As Matthias mentions in the comments, the Google Guava MapMaker does use SoftReferences:

    A ConcurrentMap builder, providing any combination of these features:

    • soft or weak keys,

    As mentioned in this thread, another JSR166y candidate:

    jsr166y.ConcurrentReferenceHashMap

    It provides an alternative concurrent reference map to the Google implementation (which relies on a background thread to evict entries)

    Edit (August 2012)

    The Google implementation uses a background thread only when timed expiration of entries is requested. In particular, it simply uses java.util.Timer, which is not so intrusive as having a separate background thread.

    Jeremy Manson recommends, for any cache, using this feature to avoid the dangers of SoftReference: http://jeremymanson.blogspot.de/2009/07/how-hotspot-decides-to-clear_07.html

    There's another implementation from Apache Commons, namely org.apache.commons.collections.map.ReferenceMap; it does not support timed removal, but it does support choosing whether keys should be compared by identity or by equality. Moreover, this implementation is not concurrent - it can be made synchronized, but that works less well under accesses from multiple threads.