Understanding the Performance of Concurrent Data Structures on Graphics Processors

Abstract

Abstract. In this paper we revisit the design of concurrent data structures -specifically queues -and examine their performance portability with regard to the move from conventional CPUs to graphics processors. We have looked at both lock-based and lock-free algorithms and have, for comparison, implemented and optimized the same algorithms on both graphics processors and multi-core CPUs. Particular interest has been paid to study the difference between the old Tesla and the new Fermi and Kepler architectures in this context. We provide a comprehensive evaluation and analysis of our implementations on all examined platforms. Our results indicate that the queues are in general performance portable, but that platform specific optimizations are possible to increase performance. The Fermi and Kepler GPUs, with optimized atomic operations, are observed to provide excellent scalability for both lock-based and lock-free queues.

Toward Concurrent Lock-Free Queues on GPUs

“…The single thread based operations tend to incur more contentions in CAS operations. As reported by Cederman et al [5], the GPU based queue operations are slower than their multi-core equivalents.…”

“…Many GPU based libraries supporting various programming primitives have been proposed (e,g, [3], [4]). On the other hand, the FIFO queue, which is one of the most fundamental data structures and has wide applications, has only attracted limited research efforts (e.g., [5], [6]). In this paper, we propose an efficient concurrent lock-free queue for GPGPU.…”

“…Misra and Chaudhuri [6] demonstrated the usage of GPUs CAS operator to implement various concurrent data structures. Cederman et al [5] implemented the concurrent lock-free queue proposed by Michael and Soctt on GPUs. Both of these works only enable a lower level of performance than what can be achieved on multi-core platforms.…”

Design and Evaluation of Scalable Concurrent Queues for Many-Core Architectures

“…Finally, several of these queues have been evaluated on CUDA GPUs by Cederman et al [3]. Out of a number of lock-based and two lock-free designs (i.e., MS-queue and TZqueue), they conclude that for higher concurrency, the two lock-free queue designs are nearly always highest performing.…”

“…By using a single thread to perform all operations on the queue at any given time, the maximum throughput at any level of contention is the same. At all numbers of threads we model the combining queue using the atomic latency for one thread in terms of Equation 3. This form is effectively the same as one would use for a serial queue, except that an additional read is performed to determine the operation to perform.…”