droid/scripts/run_preprocessing_6machines.sh

#!/bin/bash
# DROID Preprocessing: Distributed Multi-Machine Processing
# Usage: bash run_preprocessing_6machines.sh --machine-id 0 --total-machines 6

set -e

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "$SCRIPT_DIR"

# Parse arguments
MACHINE_ID=""
TOTAL_MACHINES=""
NUM_GPUS=8
OUTPUT_DIR="/mnt/kevin/data/droid_preprocessed"
MAX_FRAMES=400
PYTHON="/mnt/kevin/envs/miniconda3/envs/atm_ati_vdm_droid/bin/python"

while [[ $# -gt 0 ]]; do
    case $1 in
        --machine-id)
            MACHINE_ID="$2"
            shift 2
            ;;
        --total-machines)
            TOTAL_MACHINES="$2"
            shift 2
            ;;
        --num-gpus)
            NUM_GPUS="$2"
            shift 2
            ;;
        --output-dir)
            OUTPUT_DIR="$2"
            shift 2
            ;;
        --max-frames)
            MAX_FRAMES="$2"
            shift 2
            ;;
        --python)
            PYTHON="$2"
            shift 2
            ;;
        *)
            echo "Unknown option: $1"
            echo "Usage: $0 --machine-id <id> --total-machines <num> [--num-gpus 8] [--output-dir /tmp/droid_rlds_final] [--max-frames 400] [--python /path/to/python]"
            exit 1
            ;;
    esac
done

# Validate required arguments
if [ -z "$MACHINE_ID" ] || [ -z "$TOTAL_MACHINES" ]; then
    echo "Error: --machine-id and --total-machines are required"
    echo ""
    echo "Usage: $0 --machine-id <id> --total-machines <num> [options]"
    echo ""
    echo "Required arguments:"
    echo "  --machine-id <id>           Machine ID (0-based, e.g., 0, 1, 2, ...)"
    echo "  --total-machines <num>      Total number of machines (e.g., 6)"
    echo ""
    echo "Optional arguments:"
    echo "  --num-gpus <num>            Number of GPUs per machine (default: 8)"
    echo "  --output-dir <dir>          Output directory (default: /mnt/kevin/data/droid_preprocessed)"
    echo "  --max-frames <num>          Max frames per episode (default: 400)"
    echo "  --python <path>             Python interpreter path (default: atm_ati_vdm_droid env)"
    echo ""
    echo "Example:"
    echo "  $0 --machine-id 0 --total-machines 6"
    exit 1
fi

# Validate machine ID range
if [ "$MACHINE_ID" -lt 0 ] || [ "$MACHINE_ID" -ge "$TOTAL_MACHINES" ]; then
    echo "Error: machine-id must be between 0 and $((TOTAL_MACHINES - 1))"
    exit 1
fi

# Constants
TOTAL_EPISODES=92233  # 0-92232 inclusive

# Calculate episode range for this machine
EPISODES_PER_MACHINE=$((TOTAL_EPISODES / TOTAL_MACHINES))
REMAINDER=$((TOTAL_EPISODES % TOTAL_MACHINES))

# First REMAINDER machines get one extra episode
if [ "$MACHINE_ID" -lt "$REMAINDER" ]; then
    EPISODES_THIS_MACHINE=$((EPISODES_PER_MACHINE + 1))
    MACHINE_START=$((MACHINE_ID * EPISODES_THIS_MACHINE))
else
    EPISODES_THIS_MACHINE=$EPISODES_PER_MACHINE
    # Account for extra episodes given to earlier machines
    MACHINE_START=$((REMAINDER * (EPISODES_PER_MACHINE + 1) + (MACHINE_ID - REMAINDER) * EPISODES_PER_MACHINE))
fi

MACHINE_END=$((MACHINE_START + EPISODES_THIS_MACHINE))

# Create log directory
LOG_DIR="preprocessing_logs"
mkdir -p "$LOG_DIR"

echo "=========================================="
echo "DROID Distributed Preprocessing"
echo "=========================================="
echo "  Machine ID: $MACHINE_ID / $((TOTAL_MACHINES - 1))"
echo "  Total machines: $TOTAL_MACHINES"
echo "  Episodes for this machine: $MACHINE_START to $((MACHINE_END - 1)) ($EPISODES_THIS_MACHINE episodes)"
echo "  GPUs: $NUM_GPUS (CUDA 0-$((NUM_GPUS - 1)))"
echo "  Output: $OUTPUT_DIR"
echo "  Max frames: $MAX_FRAMES"
echo "=========================================="
echo ""

# Calculate episode distribution across GPUs
EPISODES_PER_GPU=$((EPISODES_THIS_MACHINE / NUM_GPUS))
GPU_REMAINDER=$((EPISODES_THIS_MACHINE % NUM_GPUS))

echo "Episode distribution across GPUs:"
echo "  Base episodes per GPU: $EPISODES_PER_GPU"
echo "  First $GPU_REMAINDER GPUs get +1 episode"
echo ""

# Launch GPU workers in parallel
for GPU_ID in $(seq 0 $((NUM_GPUS - 1))); do
    # First GPU_REMAINDER GPUs get one extra episode
    if [ "$GPU_ID" -lt "$GPU_REMAINDER" ]; then
        EPISODES_THIS_GPU=$((EPISODES_PER_GPU + 1))
    else
        EPISODES_THIS_GPU=$EPISODES_PER_GPU
    fi

    # Calculate cumulative start position for this GPU
    START_EP=$MACHINE_START
    for prev_gpu in $(seq 0 $((GPU_ID - 1))); do
        if [ "$prev_gpu" -lt "$GPU_REMAINDER" ]; then
            START_EP=$((START_EP + EPISODES_PER_GPU + 1))
        else
            START_EP=$((START_EP + EPISODES_PER_GPU))
        fi
    done

    END_EP=$((START_EP + EPISODES_THIS_GPU))

    # Make sure last GPU doesn't exceed machine range
    if [ $END_EP -gt $MACHINE_END ]; then
        END_EP=$MACHINE_END
    fi

    LOG_FILE="$LOG_DIR/m${MACHINE_ID}_gpu${GPU_ID}.log"

    echo "Launching GPU $GPU_ID: episodes $START_EP to $((END_EP - 1)) ($EPISODES_THIS_GPU eps) → $LOG_FILE"

    CUDA_VISIBLE_DEVICES=$GPU_ID $PYTHON preprocess_droid_rlds_final.py \
        --start-episode $START_EP \
        --end-episode $END_EP \
        --output-dir "$OUTPUT_DIR" \
        --max-frames $MAX_FRAMES \
        --gpu-id $GPU_ID \
        --save-previews 2 \
        > "$LOG_FILE" 2>&1 &
done

echo ""
echo "=========================================="
echo "All $NUM_GPUS GPU workers launched in background"
echo "=========================================="
echo ""
echo "Monitor progress:"
echo "  # Watch all GPU logs for this machine"
echo "  tail -f $LOG_DIR/m${MACHINE_ID}_gpu*.log"
echo ""
echo "  # Watch specific GPU"
echo "  tail -f $LOG_DIR/m${MACHINE_ID}_gpu0.log"
echo ""
echo "  # Check running processes"
echo "  ps aux | grep preprocess_droid_rlds_final"
echo ""
echo "  # Monitor GPU usage"
echo "  watch -n 1 nvidia-smi"
echo ""
echo "  # Count processed episodes (all machines)"
echo "  ls $OUTPUT_DIR/data/*.npz | wc -l"
echo ""
echo "Wait for all workers to finish (blocking):"
echo "  wait"
echo ""

# Wait for all background jobs to finish
wait

echo ""
echo "=========================================="
echo "Machine $MACHINE_ID: All GPU workers completed!"
echo "=========================================="
echo ""
echo "Check metadata:"
echo "  cat $OUTPUT_DIR/metadata_gpu*.json | jq '.processing_stats'"
echo ""