An alternative set of interfaces and implementations of Java collections, focused on performance
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
src/es/darkhogg/lowendcoll
.gitignore
LICENSE
README.md
build.xml

README.md

Low-End Collections

The Low-End Collections Framework is a re-implementation of most of the Java Collections Framework base classes and interfaces that focuses on high-performance and low memory allocations at the cost of some memory size and usage safety.

WARNING: Do not use this library unless you really need to. Read this README file carefully to see if you should use it or not. These classes must only be used for situations that require it, as they include some intentional safety vulnerabilities for the sake of speed and low memory allocation.

Its usage should be limited to the internals of higher-level classes or for code that knows what this library is and can use it accordingly. These collections are intentionally incompatible with regular Java Collections to avoid misuse.

Motivations

Although the Java Collections Framework is great and should be used in any situation that requires collections, they implement some code that is not well suited for performance-critical applications such as games or memory-critical environments such as Android or (again) games.

After working on some code where I needed lists and maps that were not as heavy as the regular JCF ones, I decided to implement my own collections mini-framework with a few concerns in mind:

  • Iterators should not be created constantly. On games, some collections are iterated multiple times per frame, which results in hundreds of iterations per second. If every iteration adds an object to the heap, the GC will have to run more frequently, severely impacting performance.

  • Iterators should not be safe if it impacts performance. Most iterations do a single full pass on a collection and are then forgotten. Thread safety or fail-fast are of little value in a single-threaded environment and users should take care themselves of not modifying the original collection while iterating.

  • HashMaps should be cache-friendly. The JCF implementation is open, meaning that elements are not stored on the hash table itself, but on linked lists referenced from that table. This adds objects to the heap every time a new key is inserted, and increases indirection, making cache misses more frequent. By using a closed hash table, we eliminate both problems. There's a potential increase of computational complexity for high load factors because addresses on the table are reused for other keys, but that should not be noticeable if the whole table is cached.

  • A direct implementation of Collection is necessary for those situations where you just want to keep elements in some kind of structure but order is not important and you need fast traversal and removal times.

That's how Low-End Collections was born, and those are the core principles.

Particularities

As it has been already mentioned, there are a number of particularities that you must know about when using this library. Most of the usage is equivalent to the JCF so you can feel at home.

Collection Sizes

In JCF, the size() method is defined so that it returns the number of elements in the collection or Integer.MAX_VALUE if there are more. This is not valid on LEC.

Instead, a new method is included, isFull(), that tells us if the collection is full and therefore doesn't allow any more elements. All add methods must fail if the collection is full.

This means that, in LEC, size() returns the exact number of elements in the collection and that adding elements beyond Integer.MAX_VALUE is not pemitted under any circumstances, even for structures that technically allow it.

Note however that it would make little sense to have Integer.MAX_VALUE elements, as it will use at least 235 bytes of memory just for the pointers. That's 32 GiB for just a collection of pointers. Add data to the mix and and you have an unrealistic situation where a single collection is using more than 100 GiB of memory.

Map Simplicity

In JCF, the Map interface is really complex, mainly because of the introduction of key sets, value collections and entry sets. Although they are really useful in a lot of situations, LEC does not include them.

Instead, LEC makes a LowEndMap<K, V> an subinterface of Iterable<LowEndMap.Entry<K, V>>, so maps can be iterated as if they were collections of mappings.

Entries are simplified as well, so no setValue is available. NOTE: This may change in the future, as I think setValue is a nice way of modifying a mapping as it's iterated.

Shared Iterators

In LEC, there are two ways of getting an iterator from a collection or a map: iterator() and newIterator().

The newIterator method is the equivalent of the JCF iterator: It returns a newly allocated iterator that nobody else will get, and will be GC'd when you stop using it. In the case of maps, these iterators will also return newly allocated entries.

The iterator method, however, returns a shared or cached iterator. That is, multiple invocations of iterator will keep returning the same iterator(s) again and again, properly resetted so you don't really notice or care about it. There are some implications on this approach:

  • You can't nest loops more times that iterators are cached. For example, suppose coll is a collection that keeps only one iterator.

    // Calls iterator() on coll to return a new iterator
    for (Object elem1 : coll) {
        // Calls iterator() on coll and returns *THE SAME* iterator
        for (Object elem2 : coll) {
            // Some code
        }
    }
    

    This code is supposed to iterate the collection by getting every two-element combination possible. And in fact, it would if it either was a normal JCF collection or the collection had kept more than one iterator.

    Instead, that code works as follows:

    • The first loop is entered. coll.iterator() is called.
    • The first iteration of the loop sets elem1 to the first element of the collection.
    • The second loop is entered. coll.iterator() is called, so the same iterator is returned and resetted.
    • The loop completes normally, elem2 has been set to every item in the collection.
    • The first loop completes immediately: The iterator hasNext() method returns false because the second loop had already exhausted the iterator.

    To avoid this, use newIterator() if you need to nest loops beyond the number of cached iterators. All implementations of LEC interfaces should state clearly how many loops can be nested. Standard implementations are implemented so that as much as 3 loops can be nested, so the above code would be correct if performed over a standard collection.

  • LowEndMap shared iterators also return shared entries. When you call next on a shared map iterator, the same object is always returned for that iterator. Hence, these entries should not be stored for future use, as they would always be the same and reference the last iterated mapping.

    Note that this behaviour is not the same for non-shared iterators. Iterators returned by LowEndMap#newIterator() return new entries, although they are not views of the mapping as in JCF maps.

  • Iterating with shared iterators, however, is not as bad as it looks. You get one advantage you need on memory-critical applications: There are no memory allocations derived from collection iteration, except for the first few ones. This can save you a few GC's per second.

Interfacing with JCF

Although Low-End Collections classes and interfaces do not directly implement or extend the JCF ones, there's still hope if you need to pass a LEC as a regular Java collection: the LowEndCollections utility class.

The LowEndCollections class offers a static method for every LEC interface that returns a regular JCF of the same kind. The wrapped collections have the following properties:

  • The iterator method of the collections returned by toJavaCollection, toJavaList and toJavaSet call the wrapped collection's newIterator method, hiding away the concept of shared iterators and letting clients use the collections as actual regular JCF collections.

  • Similarly, the listIterator method of the lists returned by toJavaList calls the wrapped list's newListIterator method.

  • If a LowEndList implements RandomAccess, the JCF List returned by toJavaList also implements RandomAccess.

  • Maps returned by toJavaMap implement their entrySets by creating an AbstractSet that returns as their iterator a wrapped version of the original map's newIterator.