面向RISC-V集群的指令感知容器调度框架OA

As RISC-V processors expand from embedded systems to data centers,their rich instruction set extensions offer new opportuni-ties for specialized acceleration.However,existing cloud-native schedulers typically treat CPUs as homogeneous resources,struggling to identify and leverage instruction-level heterogeneity,which limits potential performance gains.To address this,we propose a container scheduling framework for RISC-V clusters that matches application-level instruction preferences with the extension capabilities of target nodes.The framework comprises two stages:an offline analysis phase to identify runtime hotspots and infer instruction requirements,and an online scheduling phase that integrates these requirements into the Kubernetes scheduling process using native affinity mechanisms,without modifying the core scheduler.We implemented a prototype and evaluated it using representative workloads.The results demon-strate that for large language model inference,scheduling tasks to nodes supporting RVV can accelerate the context construction phase by up to 11 times.These findings indicate that combining the extensibility of the RISC-V ISA with instruction-aware scheduling can signif-icantly enhance system performance and efficiency across various scenarios.