gflow - GPU Job Scheduler

gflow is an efficient tool for scheduling and managing GPU tasks, supporting task submission from the command line and running tasks in the background. Built in Rust, it provides a simple and easy-to-use interface for single-node or distributed GPU task scheduling.

Snapshot

Key Features

• GPU Task Scheduling: Supports queuing, scheduling, and management of GPU tasks.
• Parallel Execution: Allows multiple GPU tasks to run simultaneously, maximizing GPU resource utilization.
• Command-Line Tool: Provides the CLI tool gflow for submitting tasks, and gflowd for background task scheduling.
• tmux Integration: Uses tmux to manage background tasks and track task execution status in real-time.
• TCP Submission: Submit tasks via a TCP service, making it easy to integrate with other systems.

Installation

Install via cargo (Recommended)

You can use cargo to compile and install gflow and gflowd:

cargo install gflow

Build Manually

1. Clone the repository:

   git clone https://github.com/AndPuQing/gflow.git
   cd gflow
   

2. Build the project using cargo:

   cargo build --release
   

   This will generate the gflow and gflowd executables in the target/release/ directory.

Usage

Start the Scheduler

Start the GPU task scheduler using gflow:

gflow up

Submit a Task

Submit scripts using the gflow CLI

Submit GPU tasks using the gflow command-line tool:

gflow submit test.sh --gpu 1 --conda-env myenv

• --gpu: The number of GPUs to allocate for the task.
• --conda-env: The Conda environment to activate before running the task.

Submit commands using the gflow CLI

Submit GPU tasks using the gflow command-line tool:

gflow submit "python test.py" --gpu 1 --conda-env myenv

Task Scheduling Flow

1. When submitting a task, gflow sends a TCP request to the scheduler.
2. The gflowd scheduler allocates tasks based on available GPU resources.
3. Background tasks are executed using tmux, and the scheduler monitors task status in real-time.
4. The scheduler ensures each task is executed on suitable resources and allocates GPUs in priority order.

[!WARNING] The gflow does not save task snapshots, meaning that if the associated files are deleted, the task will fail.

Configuration

gflow and gflowd provide several configuration options that you can adjust as needed:

• Configuration files: You can customize the scheduling behavior by modifying the gflowd configuration file.
• Environment variables: For example, set GFLOW_LOG_LEVEL=debug to configure the logging level.

Contributing

If you find any bugs or have feature requests, feel free to create an Issue and contribute by submitting Pull Requests.

TODO

• Support GPU task scheduling in a multi-node environment.
• Add task prioritization and resource quota management.
• Improve task retry mechanism on failure.
• Implement task result feedback and log management.

License

gflow is licensed under the MIT License. See LICENSE for more details.