A Concurrent Relaxed Queue for Unordered Parallel Accesses on GPUs
A Concurrent Relaxed Queue for Unordered Parallel Accesses on GPUs
复制标题
DOI：
10.1109/csci58124.2022.00243
复制
发表时间：
2022-12
期刊：
影响因子：
--
通讯作者：
Mengshen Zhao;David Troendle;Byunghyun Jang
中科院分区：
文献类型：
--
作者：
Mengshen Zhao;David Troendle;Byunghyun Jang
我们为图形处理单元（GPU）提出了一种宽松的并发队列，它支持无序入队和/或出队操作组。当组大小为1时，它成为传统的严格先进先出（FIFO）队列。我们将这些组称为并行操作组 (POG)，并利用持久线程模型来支持任意数量的 POG 的处理。为了最大限度地减少线程争用和同步开销，每个 POG 中的队列操作作为一个组进行处理，然后在单个更新过程中提交到队列。该实验将我们提出的队列与合成输入上的非阻塞并发队列（基线）进行比较，模拟现实应用中存在的 POG 的不同配置。实验表明，在处理大量 POG 时，我们的宽松队列比基线实现了显着的加速，同时保持了良好的可扩展性。
We propose a relaxed concurrent queue for Graphic Processing Units (GPUs) which supports groups of unordered enqueue and/or dequeue operations. When the group size is one, it becomes a conventional strict First-In-First-Out (FIFO) queue. We call these groups Parallel Operations Groups (POGs) and leverage a persistent thread model to support the processing of an arbitrary number of POGs. To minimize thread contention and synchronization overhead, queue operations in each POG are processed as a group then committed to the queue in a single update process. The experiment compares our proposed queue with a non-blocking concurrent queue (baseline) on synthetic inputs that simulate the diverse configurations of POGs present in realworld applications. The experiments show that our relaxed queue achieves a significant speed up over the baseline when processing a large number of POGs while retaining good scalability.