Cloud warehouses are increasingly adopting CPU–GPU collaborative systems to leverage diverse types and levels of parallelism in applications. These environments are shared among multiple clients to achieve maximum resource utilization with energyf efficiency and scalability. While OpenCL simplifies resource provisioning in such heterogeneous systems, ensuring the effective distribution of tasks remains challenging as CPU–GPU available architectures and workload characteristics can vary significantly. This study addresses the challenge of efficiently provisioning resources in OpenCL-based CPU–GPU cloud environments. To tackle this challenge, we introduce MultiProvision, a Design Space Exploration tool for multi-tenant resource provisioning in CPU–GPU environments. MultiProvision facilitates the identification of the most suitable provisioning strategy for a given workload and architecture scenario in a transparent manner. Through comprehensive evaluations encompassing various architecture combinations and workloads, we demonstrate that the choice of the most efficient provisioning strategy depends on the target architecture, workload characteristics, and optimization objectives, such as makespan or energy. We show that the appropriate strategy can achieve remarkable gains of up to 13.15\(\times \) in makespan and 4.52\(\times \) in energy compared to a GPU-only execution.

云仓库越来越多地采用 CPU-GPU 协作系统来利用应用程序中不同类型和级别的并行性。这些环境在多个客户端之间共享，以实现最大的资源利用率、能源效率和可扩展性。虽然 OpenCL 简化了此类异构系统中的资源配置，但确保任务的有效分配仍然具有挑战性，因为 CPU-GPU 可用架构和工作负载特征可能存在很大差异。本研究解决了在基于 OpenCL 的 CPU-GPU 云环境中高效配置资源的挑战。为了应对这一挑战，我们引入了 MultiProvision，这是一种设计空间探索工具，用于在 CPU-GPU 环境中配置多租户资源。MultiProvision 有助于以透明的方式识别给定工作负载和架构场景的最合适的配置策略。通过对各种架构组合和工作负载的综合评估，我们证明最有效的配置策略的选择取决于目标架构、工作负载特征和优化目标，例如完工时间或能源。我们表明，与仅使用 GPU 执行相比， 适当的策略可以在完工时间上实现高达 13.15 \(\times \)的显着增益，在能量方面实现高达 4.52 \(\times \) 的显着增益。