Cellular automata (CA) are discrete computational models widely used to simulate complex systems through simple, localized rules. While CA are primarily valued for their ability to visualize complex phenomena, certain simulations, such as traffic models or electrical circuits, demand high-performance processing to be practical. Being a special case of stencil computations, CA are well-suited for data parallel models and can benefit from both vectorization and GPU offloading. However, the performance optimization of CA has not been thoroughly explored, leaving many open questions. In this paper, we propose a bit-plane data encoding of CA cell states that enables efficient bitwise vectorization and works well with well-known optimizations like temporal blocking. We implemented 10 different CA using this technique to demonstrate the versatility of our approach and performed extensive evaluation. Furthermore, our implementation took advantage of Cellato abstraction, which allows simple CA rule definitions using C++ templates while abstracting away the complex implementation details. Our approach offers speedup of two orders of magnitude (comparing baseline and optimized CUDA implementations) whilst maintaining code simplicity and ease of use for the end users.
Sparse convolution plays a crucial role in applications such as autonomous driving and has been receiving increasing attention. Due to the high sparsity of the voxelized input point cloud data in sparse convolution, running sparse CNN efficiently on general hardware is a severe challenge. To accelerate the execution of sparse convolution, several dataflow strategies have emerged sequentially: gather-GEMM-scatter, fetch-on-demand, and implicit GEMM. Among these, the recently introduced implicit GEMM dataflow demonstrates high efficiency. However, it still incurs substantial redundant computation, which wastes computational resources and consequently degrades overall performance. To address redundant computations, we propose TC-SparIG, a comprehensive framework that incorporates a novel sparse format called HEB-Sparse. Using an efficient encoding and decoding scheme, HEB-Sparse directly identifies the blocks that require computation, eliminating the need to traverse ineffective blocks and check whether to skip. At a finer granularity, we design a Tensor Core level mask to reduce data movement and computation. Besides, we maintain a cache to reduce DRAM access to the mask and implement loop reordering to enhance data locality. Experimental results demonstrate that the TC-SparIG implementation of sparse convolution based on implicit GEMM dataflow can achieve an average speedup of 1.63× and 1.44× compared to state-of-the-art TorchSparse++ and SpConv 2.3.6, respectively.
This paper presents a novel adaptive machine learning (ML) framework for predicting computational tasks execution time in OpenCL environments, specifically designed to address load balancing challenges within Central Processing Unit (CPU) and Graphics Processing Unit (GPU) Heterogeneous Computing Systems (HCS). Finding a universal solution for load balancing is complex given the diverse hardware configurations and computational tasks, ranging from cybersecurity to ML workloads. The focus on CPU-GPU systems enables our solution to effectively navigate the specific demands of modern heterogeneous computing environments. The framework proposed in the paper implements a regression-based deep learning network to analyze OpenCL kernels and monitor the computing environment's state. The ML model, which is a part of the proposed framework, excels at predicting kernel execution times across various devices, enabling dynamic and responsive load balancing. The method underlying the framework operation differs from traditional static code analyzer-based approaches, which specify execution devices without considering critical runtime features extracted through kernel profiling. The framework enhances performance through dynamic selection of devices based on system status real-time monitoring. The proposed solution, integrated with a comprehensive profiling system, analyzes memory access patterns, data transfer rates, and others. The regression-based task modeling approach provides precise performance measurements, avoiding the limitations of classification methods that require redundant task executions. Empirical evaluation demonstrates substantial improvements achieving up to 9.8% enhancement in load balancing efficiency during operational fine-tuning.
Cloud is a distributed computing system that offers resources and services depending on the need. The primary concerns with managing cloud resources will be task scheduling and resource provisioning in a cloud environment. Computational complexity and the interdependency of scientific procedures restrict the use of cloud resource management. Balanced load is accomplished by emphasis on workflow detection and scheduling based on utilization percentage. A novel two stage framework is designed to reduce the number of jobs in the workflow scheduling jobs that cannot meet their deadlines. This paper proposes a new approach called Hybridized Fennec Rat Swarm Optimization (HFe-RSO), which assigns priority to the tasks. Then, a workflow scheduling model, Enhanced Soft Actor-Critic Deep Reinforcement Learning (ESAC-DRL), is developed for efficient resource management by considering the task priority and utilization rate. Workflow benchmarks such as Montage, CyberShake, Sipht, Epigenomics, and Inspiral are used to assess the performance of the proposed model. The performance of the proposed approach is analysed using cost, makespan, and utilization rate, and compared with state-of-the-art approaches. By using the proposed approach, the cost of $259.11, makespan of 99.84ms, and the utilization rate 96.20%, was achieved. The experimental results demonstrate the significance of the proposed approach over other existing approaches.
Planning unmanned aerial vehicles (UAVs) by considering the optimization concerns about the enemy threats, fuel consumption or usage and maneuver limitations has a critical importance on the task being performed and safety of the flight. The Back-and-Forth (BaF) algorithm has been introduced recently as a solving technique for the geometrical description of the UAV path planning problem and its sole performance and positive contribution on the capabilities of another methods have been validated. In this study, the greedy heuristic guided by the BaF algorithm was redesigned with the purpose of getting the power of a multi-core architecture and a new geometrical UAV path planner named as the Combined Paths for short ComPaths was proposed. The performance of the ComPaths was evaluated over 3 different battlefield scenarios and 12 test cases derived from them. Moreover, the paths found by the ComPaths were compared to the paths of other 15 heuristic or meta-heuristic backed planners. Comparative studies between the ComPaths and other tested techniques allowed to conclude that the ComPaths is among the top three best planners for all cases even though it makes nearly 43 to 49 times less function calls than the remaining competitors.
