القائمة
Knowledge base from '交我算 HPC+AI平台用户手册' by 上海交通大学网络信息中心. Use when answering questions about HPC cluster usage, SLURM job scheduling, module environment, data transfer, GPU computing, and common troubleshooting on SJTU HPC clusters.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Trigger on SJTU HPC, 交我算, Slurm, srun, sbatch, IDEKube, hpc/sy SSH; use for cluster jobs and transfer.
Log in to the SJTU HPC platform (also known as "交我算") as the user to perform job queries, submissions, cancellations, and data management. Use this skill when the user requests operations related to HPC or "交我算".
| name | sjtu-hpc |
| description | Knowledge base from '交我算 HPC+AI平台用户手册' by 上海交通大学网络信息中心. Use when answering questions about HPC cluster usage, SLURM job scheduling, module environment, data transfer, GPU computing, and common troubleshooting on SJTU HPC clusters. |
SLURM, data transfer, GPU, module, etc.; I find and read the relevant chapterch01; I load that specific chapterWhen you ask about a topic not covered in Core Frameworks below, I will read the relevant chapter file before answering.
SLURM (Simple Linux Utility for Resource Management) is the workload manager for all SJTU HPC clusters (Siyuan-1, Pi 2.0, AI, ARM).
核心命令:
sbatch <script.slurm> — 提交批处理作业squeue — 查看排队/运行中的作业 (-j <jobid>, -u <user>, --state=R)scancel <jobid> — 取消作业sinfo — 查看集群节点状态 (-N, --states=idle, --partition=<queue>)sacct — 查看已完成作业报告 (-j <jobid>, --format=..., -S YYYY-MM-DD)scontrol — 查看/修改作业参数 (show job <jobid>, hold, release)srun — 启动交互式作业 (srun -p <queue> -n <cores> --pty /bin/bash)salloc — 申请资源后登录计算节点作业状态: R(运行中), PD(排队中), CG(即将完成), CD(已完成), CANCELLED(已取消), FAILED(运行失败), TIMEOUT(超时), NODE_FAIL(节点故障)
| 集群 | 队列 | 适用场景 | 单节点配置 | 核数限制 | 最长运行时间 |
|---|---|---|---|---|---|
| 思源一号 | 64c512g | CPU通用计算 | 64核, 512G | 1~60000核 | 7天 |
| 思源一号 | a100 | GPU计算(A100) | 64核, 512G, 4×A100 40G | 1~92卡 | 7天 |
| Pi2.0 | small | 小规模测试 | 40核, 192G | 1~20核 | 7天 |
| Pi2.0 | cpu | CPU大规模并行 | 40核, 192G | 40~24000核 | 7天 |
| Pi2.0 | huge | 大内存(3T) | 80核, 3T | 1~80核 | 2天 |
| Pi2.0 | 192c6t | 超大内存(6T) | 192核, 6T | 1~192核 | 2天 |
| Pi2.0/AI | dgx2 | GPU(V100) | 96核, 1.45T, 16×V100 | 1~128卡 | 7天 |
| ARM | arm128c256g | ARM计算 | 128核, 256G | 1~12800核 | 7天 |
| 通用 | debug | 短时间测试 | — | 最多2节点 | 20~60分钟 |
选择原则: CPU大规模→cpu/64c512g; GPU→dgx2/a100; 大内存→huge/192c6t; 小测试→small/debug
集群通过 module 系统预部署数百种软件。用户无需自行安装即可加载使用。
module avail — 列出所有可用模块module load <name> — 加载模块（不加版本号加载默认版）module list — 查看已加载模块module purge — 清除所有已加载模块module unload <name> — 卸载指定模块module show <name> — 查看模块具体信息注意事项: 编译器与MPI库需匹配。加载 openmpi 模块时，若已加载 gcc 则加载 gcc 编译版，否则加载默认版。
/lustre/home/acct-xxxx/yyyysydata.hpc.sjtu.edu.cndata.hpc.sjtu.edu.cnscp -r, rsync --archive --partial --progress| # | Title | Key Topics |
|---|---|---|
| ch01 | 平台概览与硬件资源 | 集群架构, 思源一号, Pi2.0, AI, ARM, 致远一号 |
| ch02 | 账号申请与集群登录 | 主账号/子账号, SSH登录, HPC Studio, Tabby, VS Code, Tmux |
| ch03 | 文件系统与数据管理 | Lustre, GPFS, /scratch, /union, 冷存储, 数据共享 |
| ch04 | 数据传输 | data节点, scp/rsync, 集群间传输, 线下迁移, 远程挂载 |
| ch05 | SLURM作业提交与管理 | sbatch, squeue, scancel, 作业脚本, 交互式作业, 环境变量 |
| ch06 | 队列详解与资源选择 | 各队列参数, GPU配比, 作业延长, debug队列 |
| ch07 | 软件模块与编译器 | module命令, GCC, Intel, MPI库, 数学库, 软件安装 |
| ch08 | GPU计算 | A100, V100, GPU作业脚本, TensorFlow, CUDA |
| ch09 | HPC可视化平台 | HPC Studio, 远程桌面, Jupyter, RStudio, DeepSeek, MATLAB, ParaView |
| ch10 | 容器使用 | Singularity镜像构建, AI平台容器编译 |
| ch11 | 常见问题与排错 | 登录问题, 作业失败, 软件安装, 代理设置, node_fail |
| ch12 | 学科应用软件 | GROMACS, LAMMPS, VASP, OpenFOAM, WRF, Quantum ESPRESSO, 生信软件 |
| ch13 | GPU深度学习实战 | PyTorch, A100, debuga100, CUDA匹配, DDP多卡训练 |
| ch14 | 并行配置与性能优化 | BLAS线程, worker公式, 超额订阅, ProcessPoolExecutor |
| ch15 | 计费标准与成本管理 | 核时/卡时计费, 存储费用, 省钱策略, 查账充值 |
| ch16 | 实用脚本与工具 | 批量作业, 自动化, 监控, 数据管理脚本 |
| ch17 | 进阶故障排查 | 错误码, 性能诊断, 代理配置, fail2ban, 存储急救 |
| ch18 | Python/R 环境配置 | Conda管理, GPU环境, R并行, Jupyter, 环境迁移, NLP/ML工作流 |
| ch19 | HPC REST API | Token认证, SSH证书自动化, 配额查询, Agent集成 |
This skill covers the SJTU "交我算" HPC+AI platform based on the official user manual. It includes cluster architecture, SLURM job management, module environment, data transfer, GPU computing, HPC Studio visualization platform, containers, and common software applications. For topics beyond this cluster's documentation, check related skills or ask the agent directly.