The JavaTM Tutorial
Previous Page Lesson Contents Next Page Start of Tutorial > Start of Trail > Start of Lesson Search
Feedback Form

Trail: Collections
Lesson: Implementations

General Purpose Implementations

The general-purpose implementations are summarized in the table below. The table highlights their regular naming pattern: names are all of the form <Implementation> <Interface>, where <Interface> is the core collection interface implemented by the class, and <Implementation> signifies the data structure underlying the implementation.
Implementations
Hash Table Resizable Array Balanced Tree Linked List
Interfaces Set HashSet   TreeSet  
List   ArrayList   LinkedList
Map HashMap   TreeMap  
JDK 1.2 provides two implementations of each interface (with the exception of Collection (in the API reference documentation), which has no direct implementations, but serves as a least common denominator for the other collection interfaces). In each case, one implementation is clearly the primary implementation: the one to use, all other things being equal. The primary implementations are HashSet, ArrayList and HashMap. Note that the SortedSet (in the API reference documentation)and SortedMap (in the API reference documentation)interfaces do not have rows in the table above. Each of these interfaces has one implementation and these implementations (TreeSet and TreeMap) are listed in the Set and Map rows.

Not only do the implementations have consistent names, but they have consistent behavior as well. All of them implement all the optional operations contained in their interfaces. All permit null elements, keys and values. Each one is unsynchronized. All have fail-fast iterators, which detect illegal concurrent modification during iteration and fail quickly and cleanly, rather than risking arbitrary, non-deterministic behavior at an undetermined time in the future. All are Serializable, and all support a public clone method.

The fact that the new implementations are unsynchronized represents a break with the past: Vector and Hashtable, which were introduced in JDK 1.0, are synchronized. The new approach was taken because it was recognized that collections are frequently used in a manner where the synchronization is of no benefit. Such uses include single-threaded use, read-only use, and use as part of a larger data object that does its own synchronization. In general, it is good API design practice not to make users pay for a feature that they generally don't use. Further, unnecessary synchronization can result in deadlock under certain circumstances.

If you need a synchronized collection, the synchronization wrappers, described in the next section, allow any collection to be transformed into a synchronized collection. Thus, synchronization is optional for the new collection implementations where it was mandatory for the old.

As a rule of thumb, you should be thinking about the interfaces rather than the implementations. That is why there are no programming examples in this lesson. For the most part, the choice of implementation affects only performance. The preferred style, as was mentioned in the interfaces lesson, is to choose an implementation when a collection is created, and immediately assign the new collection to a variable of the corresponding interface type (or pass the collection to a method expecting an argument of the interface type). In this way, the program does not become dependent on any added methods in a given implementation, leaving the programmer or maintainer with the freedom to change implementations at the drop of a hat, if performance concerns dictate that it is the right thing to do.

The general purposes implementations are briefly discussed below. The performance of the implementations is described using words like constant, log, linear, n log(n) and quadratic. These words refer to the asymptotic upper bound on the time complexity of performing the operation. All of this is quite a mouthful, and it doesn't matter much if you don't know what it means. If you're interested, any good algorithms textbook explains this stuff. One thing to keep in mind is that this sort of performance metric has its limitations. Sometimes the nominally slower implementation may be faster for the collection size that you're actually using. When in doubt, measure the performance.

Set

The two general purpose Set (in the API reference documentation)implementations are HashSet (in the API reference documentation)and TreeSet (in the API reference documentation). It's very straightforward to decide which of these two to use. HashSet is much faster (constant time vs. log time for most operations), but offers no ordering guarantees. If you need to use the operations in the SortedSet, or in-order iteration is important to you, use TreeSet. Otherwise, use HashSet. It's a fair bet that you'll end up using HashSet most of the time.

One thing worth keeping in mind about HashSet is that iteration is linear in the sum of the number of entries and the number of buckets (the capacity). Thus, it's important to choose an appropriate initial capacity if iteration performance is important. Choosing a capacity that's too high can waste space as well as time. The default initial capacity is 101, and that's often more than you need. The initial capacity may be specified using the int constructor. To allocate a HashSet whose initial capacity is 17:

    Set s= new HashSet(17);

HashSets have one other "tuning parameter" called the load factor. If you care deeply about the space consumption of your HashSet, read the HashSet documentation for more information. Otherwise just live with the default. If you accept the default load factor but you do want to specify an initial capacity, pick a number that's about twice the size that you expect the Set to grow to. If your guess is way off, it may have to grow or you may waste a bit of space, but either way it's no big problem. If you know a prime number of about the right size, use it. If not, use an odd number. Or use an even number. It doesn't really matter much; these things might make the HashSet perform a wee bit better, but nothing to write home about.

TreeSet has no tuning parameters. With the exception of clone, neither HashSet nor TreeSet have any operations other than those required by their respective interfaces (Set and TreeSet).

List

The two general purpose List (in the API reference documentation)implementations are ArrayList (in the API reference documentation)and LinkedList (in the API reference documentation). Most of the time, you'll probably use ArrayList. It offers constant time positional access, and it's just plain fast, because it does not have to allocate a node object for each element in the List, and it can take advantage of the native method System.arraycopy when it has to move multiple elements at once. Think of ArrayList as Vector without the synchronization overhead.

If you frequently add elements to the beginning of the List, or iterate over the List deleting elements from its interior, you might want to consider LinkedList. These operations are constant time in a LinkedList but linear time in an ArrayList. But you pay a big price! Positional access is linear time in a LinkedList and constant time in an ArrayList. Furthermore, the constant factor for LinkedList is much worse. If you think that you want to use a LinkedList, measure the performance with both LinkedList and ArrayList. You may be surprised.

ArrayList has one tuning parameter, the initial capacity. It refers to the number of elements the ArrayList can hold before it has to grow. There's not much to say about it. The only ArrayList operations that are not required by List are ensureCapacity and trimToSize (which alter the excess capacity), and clone.

LinkedList has no tuning parameters, and seven optional operations, one of which is clone. The other six are addFirst, getFirst, removeFirst, addLast, getLast, and removeLast; I have very mixed feelings about them. They make it a bit more convenient to use a LinkedList as a queue or a double-ended queue (dequeue), but they prevent you from easily switching representations when you discover that ArrayList is faster.

If you need synchronization, a Vector (in the API reference documentation)will be slightly faster than an ArrayList synchronized with Collections.synchronizedList, but Vector has loads of legacy operations, so be extra careful to always manipulate the Vector with the List interface, or you'll be stuck with it for life.

If your List is fixed in size (that is, you'll never use remove, add or any of the bulk operations other than containsAll) you have a third option that's definitely worth considering. See Arrays.asList in the convenience implementations section.

Map

The two general purpose Map (in the API reference documentation)implementations are HashMap (in the API reference documentation)and TreeMap (in the API reference documentation). The situation for Map is exactly analogous to Set. If you need SortedMap operations or in-order Collection-view iteration, go for TreeMap; otherwise, go for HashMap. Everything else in the Set section also applies to Map so just re-read it.

Completeness requires that we mention Hashtable (in the API reference documentation). As with Vector and ArrayList, if you need synchronization, a Hashtable will be slightly faster than a HashMap synchronized with Collections.synchronizedMap. Again, Hashtable has loads of legacy operations, so be extra careful always to manipulate it with the Map interface, or you'll be stuck with it for life.


Previous Page Lesson Contents Next Page Start of Tutorial > Start of Trail > Start of Lesson Search
Feedback Form

Copyright 1995-2004 Sun Microsystems, Inc. All rights reserved.