sjtu-hpc

name	sjtu-hpc
description	Trigger on SJTU HPC, 交我算, Slurm, srun, sbatch, IDEKube, hpc/sy SSH; use for cluster jobs and transfer.

SJTU HPC Platform Guide (交我算)

SSH 连接安全 (fail2ban)

关键规则：

• 用别名登录: ssh hpc 或 ssh sy，不要直接用 ssh sylogin.hpc.sjtu.edu.cn（可能用错用户名）
• fail2ban 保护: 3 次认证失败 → IP 封禁 1 小时
• 失败 2 次就停: 检查 SSH 配置，不要继续尝试触发封禁
• 被封后: 换网络（手机热点）立即恢复，或等 1 小时自动解封
• 验证账号: ssh hpc-data 数据节点通常不受 fail2ban 影响，可用于验证账号/证书是否正常

推荐 SSH 配置 (~/.ssh/config):

Host hpc sy
    HostName sylogin.hpc.sjtu.edu.cn
    User YOUR_HPC_USERNAME
    IdentityFile ~/.ssh/sjtu_hpc_ed25519

Host hpc-data
    HostName sydata.hpc.sjtu.edu.cn
    User YOUR_HPC_USERNAME
    IdentityFile ~/.ssh/sjtu_hpc_ed25519

---

Quick Reference

Login & Data Nodes

Cluster	Login Node	Data Node
Siyuan-1 (思源一号)	sylogin.hpc.sjtu.edu.cn	sydata.hpc.sjtu.edu.cn
Pi 2.0 / AI	pilogin.hpc.sjtu.edu.cn	data.hpc.sjtu.edu.cn
ARM	armlogin.hpc.sjtu.edu.cn	data.hpc.sjtu.edu.cn

Queue Summary

Queue	Cluster	Max Cores	Max GPU	Max Time	Use Case
64c512g	Siyuan-1	60,000	-	7 days	Large CPU jobs
a100	Siyuan-1	-	92 cards	7 days	GPU computing (A100 40GB)
debug64c512g	Siyuan-1	128	-	60 min	CPU debugging
debuga100	Siyuan-1	64	28 virtual	20 min	GPU debug (5GB vGPU, NOT real A100)
cpu	Pi 2.0	24,000	-	7 days	CPU jobs (exclusive)
small	Pi 2.0	1-20	-	7 days	Small jobs (shared)
dgx2	AI	-	128 cards	7 days	GPU computing (V100)
huge	Pi 2.0	80	-	2 days	Large memory (3TB)

Important: Login nodes prohibit running jobs and parallel compilation. For interactive work: srun -p 64c512g -n 4 --pty /bin/bash

---

核心原则：擦干净屁股

每次操作结束后（无论正常还是异常），必须确保不留残留：

squeue -u $USER   # 检查残留作业
ls $SCRATCH        # 检查残留临时文件

违反后果：残留作业持续计费（浪费钱），资源滥用可能封号。

---

禁止操作

1. 登录节点运行计算 → 用 sbatch 或 srun -p 64c512g ...
2. 登录节点并行编译 → 申请计算节点后再 make -j
3. 登录节点大批量传输 → 使用数据传输节点
4. Debug 队列跑正式计算 → Debug 仅用于测试脚本
5. 超配额写入 → 定期清理，归档到科学大数据平台
6. 共享账号 → 每人使用自己的账号

交互式作业

srun -p 64c512g -n 4 --pty /bin/bash      # CPU 交互
srun -p a100 --gres=gpu:1 --pty /bin/bash  # GPU 交互

---

Common Commands

Command	Description
sbatch job.slurm	Submit job
squeue -u $USER	View your jobs
scancel JOBID	Cancel job
sinfo -p QUEUE	View queue status
sacct -j JOBID	View completed job info
module avail	List available modules
module load NAME	Load module

Proxy Settings (Compute Node Internet Access)

Cluster	Proxy
Siyuan-1	http://proxy2.pi.sjtu.edu.cn:3128
Pi 2.0	http://proxy.pi.sjtu.edu.cn:3004

export http_proxy=http://proxy2.pi.sjtu.edu.cn:3128   # Siyuan-1
export https_proxy=http://proxy2.pi.sjtu.edu.cn:3128

---

Detailed Documentation

Topic	Reference
SSH access & environment variables	references/access.md
Slurm job templates (CPU/GPU/Array)	references/job-templates.md
Python & R environment setup	references/python-r.md
GPU computing (A100/V100)	references/gpu.md
Data transfer methods	references/data.md
Storage architecture & quotas	references/storage.md
Queue specifications	references/queues.md
Complete Slurm commands	references/slurm-commands.md
HPC Studio web GUI	references/hpc-studio.md
Troubleshooting	references/troubleshooting.md
Billing & cost estimation	references/billing.md
Cluster monitoring dashboard	references/monitoring.md
Compute outsourcing workflow	references/compute-outsourcing.md
Parallel configuration formula	references/parallel-config.md
Utility scripts	references/utility-scripts.md

External Resources

• Documentation: https://docs.hpc.sjtu.edu.cn
• HPC Studio (Web GUI): https://studio.hpc.sjtu.edu.cn
• Account Management: https://my.hpc.sjtu.edu.cn
• Usage Monitor: https://account.hpc.sjtu.edu.cn/top/
• Cluster Status: https://mon.hpc.sjtu.edu.cn/d/Q6W7OfInk/
• Support Email: [e-mail]

Related Skills

• r-performance — 并行配置公式、BLAS 线程、future 生态
• targets-expert — 在 HPC 上跑 targets pipeline（crew.cluster）