KJCKorea
Journal Central

To address the limitations of strict FIFO lock-free queues, relaxed queues have been actively studied as a way to loosen the FIFO constraint and improve parallelism. In this work, we present the Time-Stamped Window Decoupled Queue, a novel relaxed queue that integrates time-stamping with the window-decoupled approach. Experimental results demonstrate that our design outperforms existing relaxed queues in both micro- and macro-benchmarks. Notably, in microbenchmarks, the proposed queue achieves significantly higher performance as the enqueue rate increases, surpassing other relaxed queues. In a microbenchmark with 72 threads, our queue delivers 60.67% higher throughput than the d-Choice Balanced Operations Queue.

Convolutional neural networks (CNNs) often require large receptive fields, making acceleration algorithms and GPU kernel configurations key factors in optimizing inference performance. In decomposable Winograd convolution algorithms, previous thread indexing methods lead to thread divergence, where threads within a warp are serialized due to differences in filter sizes. In this paper, we introduce a channel-wise thread indexing that eliminates thread divergence by mapping each convolutional filter channel to the same warp. This approach ensures uniform filter sizes across threads within a warp, significantly enhancing performance. Experiments show that the proposed method removes all potential thread divergence across diverse convolution configurations, including various filter sizes and input/output channel counts, reducing execution time up to 15% on state-of-the-art CNN models. These results demonstrate the potential for improving CNN computational efficiency on SIMT architectures.

Edge computing is a key technology for overcoming cloud latency and bandwidth constraints, but distributed node environments are vulnerable to both independent and correlated failures. Accordingly, this study 1) extends the existing SRN model to incorporate correlated failures, 2) defines a Capacity-Oriented Availability (COA) metric that reflects the overhead of HA and LM policies, and 3) performs sensitivity analysis through discrete event simulation. The analysis reveals that when failure intervals are long, the combined HA+LM policy performs better; however, a policy performance reversal occurs when the number of correlated failure nodes exceeds the threshold (5 nodes). This is because during large-scale failures, the node search overhead of LM delays recovery, making the HA-only policy provide higher COA. This study suggests the necessity of adaptive recovery strategies based on the scale of correlated failures.