Spring prototypes and auto-wire byType are expensive

In designing a new, very performance-sensitive part of our systems we investigated the runtime performance of retrieving beans (singleton and prototype) from a Spring bean factory versus creating them via a hand-coded factory. For the Spring code we also measured any additional overhead of auto-wiring beans and doing dependency checks on beans.

The test repeatedly retrieves 5 beans which are the roots of a highly interconnected object graph (comprising 4 other beans) from a Spring bean factory. In the prototype tests each bean in that graph is a Spring prototype bean, i.e. a new instance is create whenever a bean is needed from the bean factory. In the singleton tests each bean in that graph is a Spring singleton and so the same instance is returned every time a bean is retrieved. By comparison, the hand-coded factory always creates each object in that graph and hence behaves identical to the Spring prototype test.

The results give the time in nanoseconds for retrieving a bean (which is the root of the object graph) from the factory:

Bean retrieval / nanoseconds	Spring 2.0.5	Spring 2.0.4	Spring 2.0.3	Spring 2.0.2
Spring, prototype, autowire=byType, depend check	442365	500050	945551	958805
Spring, prototype, autowire=byName, depend check	253870	255557	667352	678782
Spring, prototype, no autowire, depend check	172412	173539	624171	639640
Spring, prototype, no autowire,no depend check	162300	162950	586144	600320
Spring, singleton, autowire=byType, depend check	608	841	1075	1134
Spring, singleton, autowire=byName,depend check	710	840	1069	1132
Spring, singleton, no autowire,depend check	609	842	1097	1161
Spring, singleton, no autowire, no depend check	614	837	1102	1135
No Spring (hand-coded factory),always create (prototype)	284	284	288	292

The numbers speak for themselves, but the shown figures visualise them.

The code we used to perform these measurements and all results are available here.