Performance Considerations

When ReentrantLock was added in Java 5.0, it offered far better contended performance than intrinsic locking. For synchronization primitives, contended performance is the key to scalability: if more resources are expended on lock management and scheduling, fewer are available for the application. A better lock implementation makes fewer system calls, forces fewer context switches, and initiates less memory-synchronization traffic on the shared memory bus, operations that are time-consuming and divert computing resources from the program.

Java 6 uses an improved algorithm for managing intrinsic locks, similar to that used by ReentrantLock, that closes the scalability gap considerably. Figure 13.1 shows the performance difference between intrinsic locks and ReentrantLock on Java 5.0 and on a prerelease build of Java 6 on a four-way Opteron system running Solaris. The curves represent the "speedup" of ReentrantLock over intrinsic locking on a single JVM version. On Java 5.0, ReentrantLock offers considerably better throughput, but on Java 6, the two are quite close.[2] The test program is the same one used in Section 11.5, this time comparing the throughput of a HashMap guarded by an intrinsic lock and by a ReentrantLock.

[2] Though this particular graph doesn't show it, the scalability difference between Java 5.0 and Java 6 really does come from improvement in intrinsic locking, rather than from regression in Reentrant-Lock.

Figure 13.1. Intrinsic Locking Versus ReentrantLock Performance on Java 5.0 and Java 6.

On Java 5.0, the performance of intrinsic locking drops dramatically in going from one thread (no contention) to more than one thread; the performance of ReentrantLock drops far less, showing its better scalability. But on Java 6, it is a different storyintrinsic locks no longer fall apart under contention, and the two scale fairly similarly.

Graphs like Figure 13.1 remind us that statements of the form "X is faster than Y" are at best short-lived. Performance and scalability are sensitive to platform factors such as CPU, processor count, cache size, and JVM characteristics, all of which can change over time. [3]

[3] When we started this book, ReentrantLock seemed the last word in lock scalability. Less than a year later, intrinsic locking gives it a good run for its money. Performance is not just a moving target, it can be a fast-moving target.

Performance is a moving target; yesterday's benchmark showing that X is faster than Y may already be out of date today.

Fairness