[tools] TACC

工具列表

/work/10397/sky1ove/ls6/
├── boltz2/.venv/
├── drugclip/.venv/
├── opendde/.venv/
├── af3_analysis/20260914/.venv/
├── tools/kalign/3.3.5/
├── af3_parameters/       # AF3 权重
├── af3_results/          # 预测结果
├── input_collection/ # 输入集合
└── databases/pdb/       # PDB 数据

• Priority：有其他任务优先安排。

• Resources：当前主要在等所需资源释放。

目录列表

目录	已用	个人配额	剩余
home	约 2.1 MiB	10 GiB	约 10 GiB
work	13.35 GiB	1 TiB	约 1,011 GiB
scratch	93.80 GiB	无个人硬限制	共享空间约 3.4 PiB

一般模型的东西会存在 work 里作为长久保存，然后输入和输出的文件都存在 scratch 里。

scratch里，访问时间超过 10 天的文件可能被清除，且 scratch 没有备份

AF3的结果存在scratch里， /scratch/10397/sky1ove/alphafold3_project/

OpenDDE 在work里

• 安装及 Python 环境：/work/10397/sky1ove/ls6/opendde

• 模型权重：/work/10397/sky1ove/ls6/opendde/data/checkpoint

输入和输出在scratch /scratch/10397/sky1ove/opendde_runs

可以把 AF3 input 的 a3m文件 (msa) 和 template 放到 work 里面统一保存。

/work/10397/sky1ove/ls6/af3_input_collection/
└── alpp_murine22/
    └── <版本或日期>/
        ├── sequences.fasta
        ├── msas/
        │   └── 原始 paired / unpaired A3M
        ├── templates/
        │   └── 原始模板 CIF
        ├── input_template.json
        └── metadata.json

ssh登录

可以 ssh <sky1ove@ls6.tacc.utexas.edu> ，然后输入密码（google password上的），和duo push里的tacc token(6 digits)，但是每次都这样会很麻烦，这里有一个办法：

Add this to your local Mac ~/.ssh/config:

Host ls6
    HostName ls6.tacc.utexas.edu
    User sky1ove

    ControlMaster auto
    ControlPath ~/.ssh/controlmasters/%r@%h:%p
    ControlPersist 8h

    ServerAliveInterval 60
    ServerAliveCountMax 3

Then create the socket directory:

mkdir -p ~/.ssh/controlmasters
chmod 700 ~/.ssh/controlmasters
chmod 600 ~/.ssh/config

Now connect with:

ssh ls6

The first connection still asks for password/MFA. After that, new ssh, scp, rsync, or sftp connections to ls6 should reuse the existing master connection and usually won’t ask MFA again until the control connection expires or dies. SSH multiplexing reuses an existing TCP/authenticated SSH session for later sessions, avoiding repeated re-authentication. Some university HPC docs explicitly recommend this to avoid retriggering MFA for repeated SSH sessions.

For your skill, I would change the default host from:

sky1ove@ls6.tacc.utexas.edu

to:

ls6

Then use:

ssh ls6
scp file.json ls6:/scratch/10397/sky1ove/alphafold3_project/
rsync -av input_json/ ls6:/scratch/10397/sky1ove/alphafold3_project/two_step/input_json/

To keep work alive after disconnecting, use tmux inside Lonestar6:

ssh ls6
tmux new -s af3

Detach:

Ctrl-b d

Reconnect later:

ssh ls6
tmux attach -t af3

To manually close the multiplexed SSH master connection:

ssh -O exit ls6

A good wording to add to your skill:

## SSH Connection Reuse

To avoid repeated password/MFA prompts during one working session, configure local SSH multiplexing for Lonestar6. This does not save passwords, MFA tokens, or TACC credentials; it only reuses an already-authenticated SSH control connection.

Recommended local `~/.ssh/config` entry:

```sshconfig
Host ls6
    HostName ls6.tacc.utexas.edu
    User sky1ove
    ControlMaster auto
    ControlPath ~/.ssh/controlmasters/%r@%h:%p
    ControlPersist 8h
    ServerAliveInterval 60
    ServerAliveCountMax 3

Create the control socket directory:

mkdir -p ~/.ssh/controlmasters
chmod 700 ~/.ssh/controlmasters
chmod 600 ~/.ssh/config

Use ssh ls6, scp ... ls6:PATH, and rsync ... ls6:PATH instead of repeatedly typing sky1ove@ls6.tacc.utexas.edu.

For long interactive work, start tmux after logging in:

tmux new -s af3

Detach with Ctrl-b d; reattach with:

tmux attach -t af3

Close the persistent SSH master manually with:

ssh -O exit ls6
For your AF3 workflow, this is especially useful because the skill repeatedly uses SSH/SFTP/SCP-like operations while creating JSONs, copying A3M backups, submitting jobs, and checking `squeue`.
::contentReference[oaicite:3]{index=3}

目录结构

work: 存放model weight的

/work/10397/sky1ove/ls6/af3_parameters/
├── af3.bin        # 1.1 GB
├── af3.bin.zst    # 974 MB
├── chunks/        # 25 MB
└── transfer_probe.txt

scratch: 放自己的文件的

/scratch/10397/sky1ove/alphafold3_project/
├── input/
│   └── input.json
├── output/
│   └── empty
├── slurm_jobs/
│   ├── af3_test.slurm
│   ├── AF3_test.o3203127
│   └── AF3_test.e3203127
└── two_step/
    ├── manifest.csv
    ├── skipped.csv
    ├── protein_input/
    │   └── HER2.json
    ├── msas/
    │   └── HER2.a3m        
    ├── input_json/
    │   └── 29 ligand JSON files
    ├── output/
    │   └── AF3 file folder
    ├── data_output/
    │   └── empty
    └── slurm_jobs/
        ├── af3_msa_HER2.slurm
        ├── af3_extract_HER2.slurm
        ├── af3_screen_20260602_HER2_yuji.slurm
        ├── extract_HER2_a3m.sh
        ├── AF3_msa_HER2.o3203601
        └── AF3_msa_HER2.e3203601

Old

Reference

Lonestar6 - TACC HPC Documentationdocs.tacc.utexas.edu/hpc/lonestar6/

login

lonestar6:

ssh username@ls6.tacc.utexas.edu

会随机给你assign一个node，你如果想要进入特定的node（1-3）

ssh username@login2.ls6.tacc.utexas.edu

进入的时候输入密码和duo push 的token

$WORK 太拥挤了，用$SCRATCH放input 和output，然后最后再把output放到$WORK

不要scp recursive copy太多文件，先压缩成.tar再传输。