The Baskets Queue

Linearizability of Persistent Memory Objects Under a Full-System-Crash Failure Model

“…Combined operations can be built together and ordered to persistence with a single pfence, then linked into the main data structure with another, reducing pfence instructions per operation. Other combining techniques (e.g., basket queues [18]) might work in a similar fashion. A transient combining array will generally result in a strictly linearizable object; leaving it persistent memory results in a durably linearizable object.…”

Cache-Aware Lock-Free Queues for Multiple Producers/Consumers and Weak Memory Consistency

“…For comparison we have also implemented the dynamic lock-free queues by Michael and Scott [8], ditto with elimination [9], the baskets queue [5], and the static cyclic array lock-free queue presented in [11]. All dynamic queues (including the new algorithm) have been implemented to support queue sizes only limited by the system's memory, i.e., using lock-free management schemes [7] or [2] and lock-free free-lists where appropriate.…”

“…However, elimination is only possible when the queue is close to be empty during the operation's invocation. Hoffman et al [5] takes another approach to increase scalability by allowing concurrent Enqueue operations to insert the new node at adjacent positions in the linked list if contention is noticed during the attempted insert at the very end of the linked list. To enable these "baskets" of concurrently inserted nodes, removed nodes are logically deleted before the actual removal from the linked list, and as the algorithm traverses through the linked list it requires stronger memory management than [7] and a strategy to avoid long chains of logically deleted nodes.…”

“…In resemblance to [8] [9][5] the new algorithm is dynamic, and in resemblance to [5] removed blocks are logically deleted, blocks are being traversed and creation of long chains are avoided. In contrast to [9][5] the new algorithm employs no special strategy …”

“…As a successful update of a shared pointer will cause a cache miss to the other threads that concurrently access that pointer, the overall strategy for improving performance and scalability of the new algorithm is to avoid accessing pointers that can be concurrently updated [5]. Moreover, our algorithm achieves fewer updates by not having shared variables with explicit information regarding which array index currently being the next active for the Enqueue or Dequeue.…”