Home

Awesome

Tool for running NN training on simple Linux cluster

TLDR

If you want to start a new W&B project on the cluster quickly, do the following

General idea

It uses SSH to run things on multiple machines without any specific cluster management software.

The project should be somewhere within your home folder (could be in a subfolder). The tool synchronizes the files of the project with all the machines in the cluster and runs specific commands on them.

It is also GPU-aware: it check which GPUs are free on which machine and runs Weights & Biases clients only on them.

It also supports fast setup of environment on all the machines in parallel.

Installation

It requires Python 3.6 or newer.

pip3 -r requirements.txt

Checkout the repository, and put the bin folder in your path.

Usage

Configuration

You need a config file. It can be called cluster.json or .cluster.json.

It should be either in your home directory (global) or your local project directory (local). You can also mix the 2: store global settings in the ~/.cluster.json and project-specific settings in the project directory.

An example global configuration file:

{
  "hosts": [
      "v02.idsia.ch",
      "v03.idsia.ch",
      "v01.idsia.ch",
      "kratos.idsia.ch",
      "minsky.idsia.ch",
      "asdf",
      "skynet.supsi.ch",
      "x01.idsia.ch",
      "x02.idsia.ch",
      "x03.idsia.ch",
      "x04.idsia.ch",
      "x05.idsia.ch",
      "nikola.idsia.ch",
      "venus.idsia.ch"
  ],
  
  "sync": {
    "exclude": [".git*"]
  },
  
  "wandb" : {
    "apikey": "YOUR_API_KEY_FROM_WANDB"
  },

  "envs": {
    "asdf": "LANG=en_US.utf-8"
  },

  "path": ["~/.local/bin"],
  "wandb_ckpt_path": "wandb/*${id}*/files/checkpoint",
  "resume_command": "--restore ${ckpt}",

  "host_config": {
    "x01.idsia.ch": {
      "user": "username",
      "key": "~/.ssh/id_rsa"
    }
  }
}

Some fields:

Example local config file

{
  "wandb" : {
    "project": "rnn_generalization_test_release"
  },

  "setup": [
    "pip3 install -U --user cython",
    "pip3 install -U --user tqdm",
    "pip3 install -U --user psutil",
    "pip3 install -U --user matplotlib",
    "pip3 install -U --user tensorboard",
    "pip3 install -U --user future",
    "pip3 install -U --user filelock",
    "pip3 install -U --user setproctitle",
    "pip3 install -U --user wandb",
    "pip3 install -U --user dataclasses",
    "pip3 install -U --user pillow",
    "pip3 install -U --user torchvision"
  ]
}

Using W&B entities

If you have your own account but also part of an organization, you might want to specify in which one to run the sweeps. In that case, just add your entity (username or organization name) in front of wandb.project in your config:

"wandb" : {
    "project": "csordas/rnn_generalization_test_release"
  },

You can always overwrite the project temporarily using the -p or --project argument.

Blacklisting a GPU

Example: add to you config

"gpu_blacklist": {
    "kratos.idsia.ch": [0,1],
    "v01.idsia.ch": [2]
}

Overriding individual commands on individual hosts

All commands are checked before it's execution against the command override list. This enables you to use the same commands on the different machines even when they need to be mapped to different things.

Example: add to your config

"commands": {
    "asdf": {
        "pip3": "~/.local/bin/pip3.6",
        "python3": "python3.6"
    },

    "kratos": {
        "pip": "pip3",
        "python": "python3"
    }
}

Specifying where to run

All commands can have a target machine, specified by argument -m. It receives a list of machines, with optional GPU specification (used just for training). To run a command on multiple machines, look at the following example:

ct -m kratos,v01 run ls

If argument -m is not used, all machines will be used.

Which GPU to use

By default the scripts autodetect free GPUs on the target machines.

Specifying GPU can overriden on per-machine basis:

ct -m 'kratos{0;1;2},v01{0-2}' run ls

It can use either a list of GPUs, separated by ; or a range of them, specified by -

Allowing to run on already used GPUs

Normally the script doesn't allow to run on GPUs that already have jobs running. However in certain situations some jobs might be small enough such that one wants to run multiple jobs on a GPU which is already used by some other process. This can be enforced by adding the -FGPU flag.

For example:

ct -m kratos -FGPU wandb sweep task.yaml

Setting up the cluster

Run ct setup in your project directory.

This will copy your SSH ID to all of the machines, so you might be asked to enter your password many times. If you don't yet have an SSH key, it will auto-generate one.

Creating a new Weights & Biases sweep

Example:

ct -m kratos wandb sweep sweeps/test.yaml

Optionally you can also specify custom sweep name:

ct -m kratos wandb sweep CoolSweep sweeps/test.yaml

In case the name is not specified, it will use the name of the yaml file instead (in the example above it will be 'test')

It will automatically synchronize your files with the target machines (see Synchronization below).

Example W&B YAML

A bit off topic, but here you go, for faster setup

program: main.py
command:
  - ${env}
  - python3
  - ${program}
  - ${args}
method: grid
metric:
  name: validation/mean_accuracy
  goal: maximize
parameters:
  log:
    value: wandb
  profile:
    distribution: categorical
    values:
      - trafo_scan
      - scan
  scan.train_split:
    distribution: categorical
    values:
      - length
      - jump
      - turn_left
  analysis.enable:
    value: 0
  stop_after:
    value: 25000
  mask_loss_weight:
    value: 3e-5
  mask_lr:
    value: 1e-2
  sweep_id_for_grid_search:
    distribution: categorical
    values:
      - 1
      - 2
      - 3
      - 4
      - 5
      - 6
      - 7
      - 8
      - 9
      - 10

Attaching new agents to an existing sweep

ct -m kratos wandb agent <sweep id>

You can copy the command from W&B dashboard and prepend ct to it.

It will automatically synchronize your files with the target machines (see Synchronization below).

Running multiple sweeps/agents on the same GPU

Append -pg <number> to your starting command (-pg = per gpu). For example:

ct -m kratos -pg 2 wandb sweep sweeps/test.yaml

Running a sweep on multiple GPUs - or multiple nodes

Use multiple GPUs on a single machine or cluster.

Append -mgpu <number> to your starting command. For example:

ct -m kratos -mgpu 2 wandb sweep sweeps/test.yaml

The jobs are started as follows:

Running multiple Weights & Biases agents

Use multiple parallel agents. It allocates free GPUs for all of them and run this many in parallel.

Append -n_runs <number> or -r <number> to your starting command. For example: If you do not specify this, the number of runs will be automatically set to cover all variants yaml file.

ct -m kratos -r 4 wandb sweep sweeps/test.yaml

Resuming Weights & Biases sweep

So far this works only with SLURM.

ct -s -m daint wandb resume idsia/lm/3fp2dk2a -mgpu 32

It will resume all the crashed experiments. If you want to resume all, even when flagged "finished", add argument -f.

Specifying how many configurations a W&B client can run

Use argument -c <number>

For example if you want to run 10 trainings with 4 parallel agents each of them running on 2 GPUs (which means 2 different nodes on daint):

ct -s -m daint -r 4 -c 10 -mgpu 2 wandb sweep sweeps/test.yaml

It will use 8 gpus in parallel.

Cleaning up Weights & Biases files

Weights & Biases stores all files both locally, and uploaded to the remote server. The local files can take up a lot of space. You can easily clear the files by the following command:

ct -m kratos wandb cleanup

Note: This command is safe to run even while sweeps are running on the target machine. It will check for the running sweeps and will not remove their folders in the W&B cache.

By default, the command assumes that W&B directory is relative to the current dir, named wandb. You can specify alternative paths with an extra argument:

ct -m kratos wandb cleanup ~/wandb

Trying to fix W&B crashed runs

Weights & Biases runs sometime show "crashed" status. If you check the log, there is no errors, it just stops at some point. This is due to an internal crash to the syncharonizaton mechansim inside the W&B agent. It can also happen if the iternet stops working for a while. In this case all the logs are stored in a local directory and they can be synchronized with the server. Run

ct wandb sync_crashed

to attempt to fix them. In my experience it works 99% of the time.

It is also possible to specify which sweep you want to synchronize by

ct wandb sync_crashed <sweep id or name>

where <sweep id or name> can be either the sweep id (the 8 random characters identifying the sweep), or the user readable name shown in the list of sweeps.

Synchronization with the local machine

ct -m kratos,v01 copy

It will copy your current working directory to all the target machines. It uses rsync, so only the modified files will be transmitted. This ensures that your code on the target machine is in perfect sync with your local one.

Synchronizing additional files

You might store some files outside of the project directory. You can add them explicitly to the sychronization list, by adding to your config:

"sync" : {
  "extra": ["~/pretrained"]
},

Gathering

You can copy back output directories from multiple machines (like Synchronization, just the other way around.)

ct -m kratos,v01 gather output

This will download and merge content of the output folder from the listed machines (into the local output folder).

If the name of files might be in conflict (the same name on multiple machines), then you can use argument -pf to the host name as a prefix to them.

Running a command

ct -m kratos,v01 run 'ls -l'

Running a command as root

ct -m kratos,v01 sudo whoami

Running a command in screen

ct -m kratos,v01 screen run 'python3 main.py'

Listing PGIDs of phantom processes

Sometimes processes can get stuck on the GPUs and using memory. In these cases nvidia-smi doesn't show them, but nvidia-smi --query-compute-apps=pid,name --format=csv does. But the PIDs are invalid. It turns out that these processes usually don't have parent processes, which can be used to detect them. You can list the PGIDs of these processes by ct list_phantom. You can also use it in conjunction with -m.

For example: ct -m kratos list_phantom

Killing phantom processes

See section on "Listing PGIDs of phantom processes" for more details on what phantom processes are.

ct -m kratos kill_phantom

Note: Killing other user's processes requires sudo.

Debugging

If something doesn't work, try adding argument -d. It will display all bash commands it used to do the specific task. By running them (or parts of them) manually can help you figure out what is the problem.

For example: ct -m kratos,v01 -d wandb agent <run id>

SLURM support

Commands supported on SLURM clusters works exactly like the locals, except that -s/--slurm switch should be passed such that no command is run accidentally on the cluster. Currently supported commands are copy, wandb sweep and wandb agent. Expected duration could also be passed to wandb commands in the form of -t hh:mm:ss (detaults to 23:59:00).

For example to run each configuration in a sweep on a node run:

ct -s -m daint wandb sweep sweep.yaml

For example to run a sweep on 20 nodes for 10 hours:

ct -s -m daint wandb sweep sweep.yaml -r 20 -t 10:00:00

In order for SLURM to work, it needs additional entries in the cluster.json. The SLURM head node should not be listed under the "hosts" array, but under a separate "slurm" dict. For example:

"slurm": {
  "daint": {
    "target_dir": "$SCRATCH",
    "modules": ["daint-gpu", "PyExtensions", "PyTorch"],
    "account": "your_account",
    "cscs_auth": {
      "username": "<username>",
      "password": "<password>",
      "otp_secret": "<otp secret>"
    }
  }
}

Alternative example, with multiple machines with different GPU types:

"slurm": {
  "sc": {
    "template": "stanford",
    "target_dir": "/some/common/path/slurm",
    "machines":[
      "bigmachine",
      "smallmachine"
    ],

    "default_partition": "standard",
    "default_cpus_per_gpu": 8,
    "default_mem_per_gpu": 16,

    "partition_map": {
      "standard": "small,big-standard",
      "high": "small-hi,big-hi"
    },
    
    "gpu_map":{
      "titanxp": ["smallmachine"],
      "a100": ["bigmachine"]
    }
  }
},

Here, headnode is the hostname for the node which is used to SLURM the slurm commands, and bigmachine and smallmachine are the names of individual machines in SLURM.

Obligatory arguments (separately for each target):

Optional arguments:

Accessing SLURM behind a front-end server

If you can't directly access the SLURM server, but it is behind another front-end server, you can add a similar entry to your local ~/.ssh/config:

Host ela
   Hostname ela.cscs.ch
   User <username>
   IdentityFile ~/.ssh/id_rsa_cscs
Host daint
   Hostname daint.cscs.ch
   User <username>
   IdentityFile ~/.ssh/id_rsa_cscs
   ProxyJump ela

Here the ProxyJump line makes all connections going to daint go through ela.

Note that in this case you should not add the full url the config.json, but the name used after Host in the ~/.ssh/config file (in this case just daint).

Getting started

It's recommended to use W&B for our server experiments. For that, you will need to set up clustertool with your W&B API key. Go to https://app.wandb.ai/settings, scroll to "API keys", click "New key" and copy your key in the config file. I recommend to put it in your global config file (~/.cluster.json), but it also works locally.

Create a local config file in your project folder, called cluster.json, as follows:

{
  "wandb" : {
    "project": "lr_tuning_test"
  }
}

Now you should setup your cluster:

ct setup

You have to do this only once, or when python version on the hosts are updated, or the setup section of the config file is changed.

In your training script, you want to log things to W&B. Import 'wandb' and initialize it as follows:

wandb.init()

Log your loss with wandb.log({"loss": loss}) periodically. Finally call wandb.join() when your training terminates.

To tune for example the learning rate, use argument parser, and create a new argument "lr". Parse and use this learning rate in your optimizer:

import argparse
args = argparse.ArgumentParser()
args.add_argument("-lr", "--lr", type=float)

opt = args.parse_args()

...

optim = torch.optim.Adam(model.parameters(), lr=opt.lr)

Start a W&B sweep as follows. First create a yaml config file:

program: main.py
command:
  - ${env}
  - python3
  - ${program}
  - ${args}
method: bayes
metric:
  name: loss
  goal: minimize
parameters:
  lr:
    min: 0.001
    max: 0.1

Save it as lr_tuning.yaml.

Next run 100 iterations of the training:

ct -m kratos,v01 -c 100 wandb sweep lr_tuning.yaml

Go to https://app.wandb.ai/, open your sweeps, and you should see your runs there.

Stopping your runs

Go to W&B sweeps page. If you don't need the data, delete the sweep. It will stop all agents immediately.

If you need your data, go to the sweep, go to the sweep config on the left side menu, and cancel the sweep.

If it still doesn't stop, you can always use ct run 'killall -9 wandb' followed by ct run 'killall -9 python3' which will terminate the runs for sure. (Note: the order is important, otherwise W&B will start new runs immediately).

Using it with conda

Add your conda bin path to either path or paths.