Objects need to be created before they can be used, and garbage-collected when they are finished with. The more objects you use, the heavier this garbage-cycling impact becomes. General object-creation statistics are actually quite difficult to measure decisively, since you must decide exactly what to measure, what size to pregrow the heap space to, how much garbage collection impacts the creation process if you let it kick in, etc.
For example, on a medium Pentium II, with heap space pregrown so that garbage collection does not have to kick in, you can get around half a million to a million simple objects created per second. If the objects are very simple, even more can be garbage-collected in one second. On the other hand, if the objects are complex, with references to other objects, and include arrays (like Vector and StringBuffer) and nonminimal constructors, the statistics plummet to less than a quarter of a million created per second, and garbage collection can drop way down to below 100,000 objects per second. Each object creation is roughly as expensive as a malloc in C, or a new in C++, and there is no easy way of creating many objects together, so you cannot take advantage of efficiencies you get using bulk allocation.
There are already runtime systems that use generational garbage collection, minimize object-creation overhead, and optimize native-code compilation. By doing this they reach up to three million objects created and collected per second (on a Pentium II), and it is likely that the average Java system should improve to get closer to that kind of performance over time. But these figures are for basic tests, optimized to show the maximum possible object-creation throughput. In a normal application with varying size objects and constructor chains, these sorts of figures cannot be obtained or even approached. Also bear in mind that you are doing nothing else in these tests apart from creating objects. In most applications, you are doing something with all those objects, making everything much slower but significantly more useful. Avoidable object creation is definitely a significant overhead for most applications, and you can easily run through millions of temporary objects using inefficient algorithms that create too many objects. In Chapter 5, we look at an example that uses the StreamTokenizer class. This class creates and dereferences a huge number of objects while it parses a stream, and the effect is to slow down processing to a crawl. The example in Chapter 5 presents a simple alternative to using a StreamTokenizer, which is 100 times faster: a large percentage of the speedup is gained from avoiding cycling through objects.