引用pbalcer · 2022-06
“Many companies are already figuring out how to best deal with the more complex memory topologies that CXL will enable [4, 5].”
举例说明许多公司正在应对 CXL 带来的更复杂内存拓扑
用 CXL 内存池和机器学习预测分配,降低云 DRAM 成本
公有云希望在满足严格性能要求的同时降低硬件成本,内存池化有望提高 DRAM 利用率,但在云性能要求下很难做到。论文提出 Pond,首个既满足云性能目标又显著降低 DRAM 成本的内存池系统,基于 CXL 标准访问池化内存。其两点关键认识是:对云生产轨迹的分析表明跨 8–16 个 socket 池化就能获得大部分收益,从而可用低延迟的小池设计;可以训练机器学习模型准确预测应给虚拟机分配多少本地内存和池内存以接近同 NUMA 节点的性能。在 158 个工作负载的评估中,Pond 把 DRAM 成本降低 7%,性能与同 NUMA 节点分配相差 1–5%。
HN 评论把它作为 CXL 内存池化的实例,说明业界已在设法应对 CXL 带来的更复杂内存拓扑,并在保持性能的同时降低成本。
“Many companies are already figuring out how to best deal with the more complex memory topologies that CXL will enable [4, 5].”
“Pond [1] which is a memory pooling system, claims to achieve the desired performance and at the same time lowering costs as one example.”