"The primary source of serialization in programs is the exclusive resource lock. The longer locks are held, the worse it gets.
Even when tasks consist only of thread-local computation, there is still serialization inherent in task dispatching," Goetz explained. "Accessing the task queue and the results container invariably involves serialization."

One answer to this problem is to replace those data structures with ones that impose less serialization. For instance, the synchronous queue has been replaced in Java 6 by a better implementation for concurrency.

The other issue is how to hold locks for less time. Goetz suggested moving thread-local computation out of synchronization blocks, "but don’t make them so small as to split atomic operations." Another trick is to replace synchronization counters with AtomicInteger.

Finally, lock splitting can reduce contention: "You often write one very large lock, and sometimes that big fat lock on your hash map is the bottleneck. You can improve things by replacing with finer granularity locks; replace a synchronized Map with ConcurrentHashMap, which has separate locks for each access."