Visible to Intel only — GUID: GUID-B554EFB3-C4F7-4BDF-9D06-0C72C4836D17
Package Contents
Parallelizing Simple Loops
Parallelizing Complex Loops
Parallelizing Data Flow and Dependence Graphs
Work Isolation
Exceptions and Cancellation
Floating-point Settings
Containers
Mutual Exclusion
Timing
Memory Allocation
The Task Scheduler
Design Patterns
Migrating from Threading Building Blocks (TBB)
Constrained APIs
Appendix A Costs of Time Slicing
Appendix B Mixing With Other Threading Packages
References
parallel_for_each Body semantics and requirements
parallel_sort ranges interface extension
TBB_malloc_replacement_log Function
Type-specified message keys for join_node
Scalable Memory Pools
Helper Functions for Expressing Graphs
concurrent_lru_cache
task_group extensions
The customizing mutex type for concurrent_hash_map
Visible to Intel only — GUID: GUID-B554EFB3-C4F7-4BDF-9D06-0C72C4836D17
Reference Counting
Problem
Destroy an object when it will no longer be used.
Context
Often it is desirable to destroy an object when it is known that it will not be used in the future. Reference counting is a common serial solution that extends to parallel programming if done carefully.
Forces
If there are cycles of references, basic reference counting is insufficient unless the cycle is explicitly broken.
Atomic counting is relatively expensive in hardware.
Solution
Thread-safe reference counting is like serial reference counting, except that the increment/decrement is done atomically, and the decrement and test “count is zero?” must act as a single atomic operation. The following example uses std::atomic<int> to achieve this.
template<typename T>
class counted {
std::atomic<int> my_count;
T value;
public:
// Construct object with a single reference to it.
counted() {my_count=1;}
// Add reference
void add_ref() {++my_count;}
// Remove reference. Return true if it was the last reference.
bool remove_ref() {return --my_count==0;}
// Get reference to underlying object
T& get() {
assert(my_count>0);
return my_value;
}
};
It is incorrect to use a separate read for testing if the count is zero. The following code would be an incorrect implementation of method remove_ref() because two threads might both execute the decrement, and then both read my_count as zero. Hence two callers would both be told incorrectly that they had removed the last reference.
--my_count; return my_count==0. // WRONG!
The decrement may need to have a release fence so that any pending writes complete before the object is deleted.
There is no simple way to atomically copy a pointer and increment its reference count, because there will be a timing hole between the copying and the increment where the reference count is too low, and thus another thread might decrement the count to zero and delete the object. Two ways to address the problem are “hazard pointers” and “pass the buck”. See the references below for details.
Variations
Atomic increment/decrement can be more than an order of magnitude more expensive than ordinary increment/decrement. The serial optimization of eliminating redundant increment/decrement operations becomes more important with atomic reference counts.
Weighted reference counting can be used to reduce costs if the pointers are unshared but the referent is shared. Associate a weight with each pointer. The reference count is the sum of the weights. A pointer x can be copied as a pointer x' without updating the reference count by splitting the original weight between x and x'. If the weight of x is too low to split, then first add a constant W to the reference count and weight of x.
References
D. Bacon and V.T. Rajan, “Concurrent Cycle Collection in Reference Counted Systems” in Proc. European Conf. on Object-Oriented Programming (June 2001). Describes a garbage collector based on reference counting that does collect cycles.
M. Michael, “Hazard Pointers: Safe Memory Reclamation for Lock-Free Objects” in IEEE Transactions on Parallel and Distributed Systems (June 2004). Describes the “hazard pointer” technique.
M. Herlihy, V. Luchangco, and M. Moir, “The Repeat Offender Problem: A Mechanism for Supporting Dynamic-Sized, Lock-Free Data Structures” in Proceedings of the 16th International Symposium on Distributed Computing (Oct. 2002). Describes the “pass the buck” technique.