TeraSort 案例SMT ON 性能分析结论:TeraSort 在本集群开 SMT 慢~2.6%,机理明确(共享执行资源争抢调度抖动中断放大),与CPU 密集高 IPC 负载应关 SMT的业界共识一致;Cloudera开 HT的建议面向 I/O 密集的通用 MR,不适用于此场景。 Sources:-AMD EPYC9005服务器 BIOS工作负载调优指南(https://m.zhiding.cn/article/3161530.htm)-SMT/HT in HPC — CloudHPCNews(https://cloudhpc.news/smt-ht-in-hpc/)-SC17:Increasing Throughput of Multiprogram HPCWorkloads(SMT Co-scheduling)(https://sc17.supercomputing.org/SC17%20Archive/tech_poster/poster_files/post181s2-file3.pdf)-AMReX nosmt 变更记录(https://archive.softwareheritage.org/browse/snapshot/e5a963a3e118bb2339e4aa4797afe3ac6ff07b1a/log/?release23.12)-Cloudera:Number of coresandmultithreading(https://docs.cloudera.com/cdp-reference-architectures/latest/cdp-pvc-base-ra/topics/ra-cdpdc-number-of-cores-multithreading.html)-Stack Overflow:nosmt 配置讨论(https://stackoverflow.com/feeds/question/78214869)上下文切换分析
