Posts

Showing posts with the label determinism

Memory management myths: determinism

Although the vast majority of programmers have now migrated to garbage collected languages and will probably never go back, there are still a few clinging to manual memory management. In most cases, the continued use of manual memory management is for good reason but some of these people are perpetuating myths in an attempt to justify avoiding garbage collection. Determinism can be a genuinely good reason to stick with manual memory management and is practically important in memory-constrained embedded devices. However, C++ programs are not as deterministic as people sometimes claim and, in particular, thread-safe reference counting using  shared_ptr  is non-deterministic. Specifically, threads holding references to shared reference-counted objects race to perform the final decrement and the thread that wins the race is responsible for destruction.

How do reference counting and tracing garbage collection compare?

Reference counting works by counting the number of references to each object. When the reference count drops to zero the object is definitely unreachable and can be recycled. The first advantage is simplicity. The second advantage is that decrements can occur as locals fall out of scope so collection can be deterministic and predictable. The third advantage is that the asymptotic complexity of reference counting does not depend upon the total size of the heap. The first problem is that cycles keep reference counts above zero even when they are unreachable, so reference counting alone cannot handle the general case. The second problem with reference counting is that incrementing and decrementing a counter in an object every time a reference to it is copied or dropped is very computationally expensive because it incurs a cache miss even if nothing else in the object is read or written. The third problem is that multithreaded reference counting is non-deterministic because increments and ...