Resource Configuration

Resource Configuration

Overview

Latch SDK provides flexible resource configuration for workflow tasks, enabling efficient execution on appropriate compute infrastructure including CPU, GPU, and memory-optimized instances.

Task Resource Decorators

Standard Decorators

The SDK provides pre-configured task decorators for common resource requirements:

@small_task

Default configuration for lightweight tasks:

from latch import small_task

@small_task
def lightweight_processing():
    """Minimal resource requirements"""
    pass

Use cases:

• File parsing and manipulation
• Simple data transformations
• Quick QC checks
• Metadata operations

@large_task

Increased CPU and memory for intensive computations:

from latch import large_task

@large_task
def intensive_computation():
    """Higher CPU and memory allocation"""
    pass

Use cases:

• Large file processing
• Complex statistical analyses
• Assembly tasks
• Multi-threaded operations

@small_gpu_task

GPU-enabled with minimal resources:

from latch import small_gpu_task

@small_gpu_task
def gpu_inference():
    """GPU-enabled task with basic resources"""
    pass

Use cases:

• Neural network inference
• Small-scale ML predictions
• GPU-accelerated libraries

@large_gpu_task

GPU-enabled with maximum resources:

from latch import large_gpu_task

@large_gpu_task
def gpu_training():
    """GPU with maximum CPU and memory"""
    pass

Use cases:

• Deep learning model training
• Protein structure prediction (AlphaFold)
• Large-scale GPU computations

Custom Task Configuration

For precise control, use the @custom_task decorator:

from latch import custom_task
from latch.resources.tasks import TaskResources

@custom_task(
    cpu=8,
    memory=32,  # GB
    storage_gib=100,
    timeout=3600,  # seconds
)
def custom_processing():
    """Task with custom resource specifications"""
    pass

Custom Task Parameters

• cpu: Number of CPU cores (integer)
• memory: Memory in GB (integer)
• storage_gib: Ephemeral storage in GiB (integer)
• timeout: Maximum execution time in seconds (integer)
• gpu: Number of GPUs (integer, 0 for CPU-only)
• gpu_type: Specific GPU model (string, e.g., "nvidia-tesla-v100")

Advanced Custom Configuration

from latch.resources.tasks import TaskResources

@custom_task(
    cpu=16,
    memory=64,
    storage_gib=500,
    timeout=7200,
    gpu=1,
    gpu_type="nvidia-tesla-a100"
)
def alphafold_prediction():
    """AlphaFold with A100 GPU and high memory"""
    pass

GPU Configuration

GPU Types

Available GPU options:

• nvidia-tesla-k80: Basic GPU for testing
• nvidia-tesla-v100: High-performance for training
• nvidia-tesla-a100: Latest generation for maximum performance

Multi-GPU Tasks

from latch import custom_task

@custom_task(
    cpu=32,
    memory=128,
    gpu=4,
    gpu_type="nvidia-tesla-v100"
)
def multi_gpu_training():
    """Distributed training across multiple GPUs"""
    pass

Resource Selection Strategies

By Computational Requirements

Memory-Intensive Tasks:

@custom_task(cpu=4, memory=128)  # High memory, moderate CPU
def genome_assembly():
    pass

CPU-Intensive Tasks:

@custom_task(cpu=64, memory=32)  # High CPU, moderate memory
def parallel_alignment():
    pass

I/O-Intensive Tasks:

@custom_task(cpu=8, memory=16, storage_gib=1000)  # Large ephemeral storage
def data_preprocessing():
    pass

By Workflow Phase

Quick Validation:

@small_task
def validate_inputs():
    """Fast input validation"""
    pass

Main Computation:

@large_task
def primary_analysis():
    """Resource-intensive analysis"""
    pass

Result Aggregation:

@small_task
def aggregate_results():
    """Lightweight result compilation"""
    pass

Workflow Resource Planning

Complete Pipeline Example

from latch import workflow, small_task, large_task, large_gpu_task
from latch.types import LatchFile

@small_task
def quality_control(fastq: LatchFile) -> LatchFile:
    """QC doesn't need much resources"""
    return qc_output

@large_task
def alignment(fastq: LatchFile) -> LatchFile:
    """Alignment benefits from more CPU"""
    return bam_output

@large_gpu_task
def variant_calling(bam: LatchFile) -> LatchFile:
    """GPU-accelerated variant caller"""
    return vcf_output

@small_task
def generate_report(vcf: LatchFile) -> LatchFile:
    """Simple report generation"""
    return report

@workflow
def genomics_pipeline(input_fastq: LatchFile) -> LatchFile:
    """Resource-optimized genomics pipeline"""
    qc = quality_control(fastq=input_fastq)
    aligned = alignment(fastq=qc)
    variants = variant_calling(bam=aligned)
    return generate_report(vcf=variants)

Timeout Configuration

Setting Timeouts

from latch import custom_task

@custom_task(
    cpu=8,
    memory=32,
    timeout=10800  # 3 hours in seconds
)
def long_running_analysis():
    """Analysis with extended timeout"""
    pass

Timeout Best Practices

1. Estimate conservatively: Add buffer time beyond expected duration
2. Monitor actual runtimes: Adjust based on real execution data
3. Default timeout: Most tasks have 1-hour default
4. Maximum timeout: Check platform limits for very long jobs

Storage Configuration

Ephemeral Storage

Configure temporary storage for intermediate files:

@custom_task(
    cpu=8,
    memory=32,
    storage_gib=500  # 500 GB temporary storage
)
def process_large_dataset():
    """Task with large intermediate files"""
    # Ephemeral storage available at /tmp
    temp_file = "/tmp/intermediate_data.bam"
    pass

Storage Guidelines

• Default storage is typically sufficient for most tasks
• Specify larger storage for tasks with large intermediate files
• Ephemeral storage is cleared after task completion
• Use LatchDir for persistent storage needs

Cost Optimization

Resource Efficiency Tips

1. Right-size resources: Don't over-allocate
2. Use appropriate decorators: Start with standard decorators
3. GPU only when needed: GPU tasks cost more
4. Parallel small tasks: Better than one large task
5. Monitor usage: Review actual resource utilization

Example: Cost-Effective Design

# INEFFICIENT: All tasks use large resources
@large_task
def validate_input():  # Over-provisioned
    pass

@large_task
def simple_transformation():  # Over-provisioned
    pass

# EFFICIENT: Right-sized resources
@small_task
def validate_input():  # Appropriate
    pass

@small_task
def simple_transformation():  # Appropriate
    pass

@large_task
def intensive_analysis():  # Appropriate
    pass

Monitoring and Debugging

Resource Usage Monitoring

During workflow execution, monitor:

• CPU utilization
• Memory usage
• GPU utilization (if applicable)
• Execution duration
• Storage consumption

Common Resource Issues

Out of Memory (OOM):

# Solution: Increase memory allocation
@custom_task(cpu=8, memory=64)  # Increased from 32 to 64 GB
def memory_intensive_task():
    pass

Timeout:

# Solution: Increase timeout
@custom_task(cpu=8, memory=32, timeout=14400)  # 4 hours
def long_task():
    pass

Insufficient Storage:

# Solution: Increase ephemeral storage
@custom_task(cpu=8, memory=32, storage_gib=1000)
def large_intermediate_files():
    pass

Conditional Resources

Dynamically allocate resources based on input:

from latch import workflow, custom_task
from latch.types import LatchFile

def get_resource_config(file_size_gb: float):
    """Determine resources based on file size"""
    if file_size_gb < 10:
        return {"cpu": 4, "memory": 16}
    elif file_size_gb < 100:
        return {"cpu": 16, "memory": 64}
    else:
        return {"cpu": 32, "memory": 128}

# Note: Resource decorators must be static
# Use multiple task variants for different sizes
@custom_task(cpu=4, memory=16)
def process_small(file: LatchFile) -> LatchFile:
    pass

@custom_task(cpu=16, memory=64)
def process_medium(file: LatchFile) -> LatchFile:
    pass

@custom_task(cpu=32, memory=128)
def process_large(file: LatchFile) -> LatchFile:
    pass

Best Practices Summary

1. Start small: Begin with standard decorators, scale up if needed
2. Profile first: Run test executions to determine actual needs
3. GPU sparingly: Only use GPU when algorithms support it
4. Parallel design: Break into smaller tasks when possible
5. Monitor and adjust: Review execution metrics and optimize
6. Document requirements: Comment why specific resources are needed
7. Test locally: Use Docker locally to validate before registration
8. Consider cost: Balance performance with cost efficiency

Platform-Specific Considerations

Available Resources

Latch platform provides:

• CPU instances: Up to 96 cores
• Memory: Up to 768 GB
• GPUs: K80, V100, A100 variants
• Storage: Configurable ephemeral storage

Resource Limits

Check current platform limits:

• Maximum CPUs per task
• Maximum memory per task
• Maximum GPU allocation
• Maximum concurrent tasks

Quotas and Limits

Be aware of workspace quotas:

• Total concurrent executions
• Total resource allocation
• Storage limits
• Execution time limits

Contact Latch support for quota increases if needed.