Running Workloads in Velda¶

This guide explains how to run workloads in Velda using the vrun command, including running on different resource pools, reattaching to sessions, and connecting to open ports.

The `vrun` Command¶

vrun is Velda's primary command for executing workloads on cloud resources. Prefix any command with vrun to run it on your Velda Infrastructure after connected to your Velda instance.

vrun will run the command on a different node with different compute setup, while keep the environment (including files, packages, environment variables) same as your instance.

Basic Syntax¶

vrun [OPTIONS] COMMAND [ARGS...]

Resource Pools¶

Resource pools define pre-configured compute profiles. Use the -P flag to specify which pool to use.

The available pool configurations vary by cluster.

Using Resource Pools¶

# List available pools
velda pool list

# All pool names below are for reference only.

# Run CPU-intensive compilation
vrun -P cpu-large make -j 16

# Run GPU training
vrun -P gpu-t4 python train.py

# High-memory data processing
vrun -P mem-xlarge python process_large_dataset.py

# Multi-GPU distributed training
vrun -P gpu-4xa100 torchrun --nproc_per_node=4 train.py

Default Pool¶

If you don't specify a pool, vrun uses the default shell pool, which typically provides a single CPU instance.

# Runs on default pool
vrun echo "Hello World"

Working with Sessions¶

Sessions allow you to run multiple commands in the same physical VM and reattach to running workloads.

Creating Named Sessions¶

Use the -s flag to assign a custom session name:

# Start a workload in a named session
vrun -s training -P gpu-a100 python train.py

Reattaching to Sessions¶

Run additional commands in the same session from any terminal:

# Check GPU usage in the training session
vrun -s training nvidia-smi

# Open an interactive shell
vrun -s training bash

Session Characteristics¶

Commands in the same session run on the same VM
Multiple commands share allocated resources
Sessions terminates when disconnected