Parallel Execution

Harbor’s parallel execution system allows you to run hundreds or thousands of trials simultaneously. This guide covers strategies for optimizing throughput, managing resources, and troubleshooting common issues.

Execution Architecture

Harbor uses an orchestrator pattern to manage parallel trial execution:

Job
 ├── Orchestrator (manages concurrency)
 ├── Trial 1 (task-001, agent-A, attempt-1)
 ├── Trial 2 (task-001, agent-A, attempt-2)
 ├── Trial 3 (task-002, agent-A, attempt-1)
 └── Trial N ...

Each trial runs independently with its own:

Environment (container)
Agent instance
Verifier
Logs directory

Orchestrator Types

Local Orchestrator

Runs trials in parallel on your local machine using asyncio:

harbor run --dataset terminal-bench@2.0 \
  --agent claude-code \
  --model anthropic/claude-opus-4-1 \
  --orchestrator local \
  --n-concurrent 8

Best for:

Development and testing
Small to medium evaluations (1-50 concurrent trials)
Controlled resource usage

Configuration:

{
  "orchestrator": {
    "type": "local",
    "n_concurrent_trials": 8,
    "quiet": false,
    "retry": {
      "max_retries": 2,
      "wait_multiplier": 1.0,
      "min_wait_sec": 1.0,
      "max_wait_sec": 60.0
    }
  }
}

Choosing Concurrency Level

The optimal concurrency depends on several factors:

CPU-Bound Tasks

For tasks that use significant CPU:

# Conservative: 1x CPU cores
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent $(nproc)

# Moderate: 2x CPU cores (if I/O wait is high)
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent $(($(nproc) * 2))

Memory-Bound Tasks

Calculate based on available memory:

# If each task needs 4GB and you have 32GB:
# 32GB total - 8GB for system = 24GB available
# 24GB / 4GB per task = 6 concurrent tasks
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent 6

I/O-Bound Tasks

For tasks that wait on API calls or network:

# Can go higher since trials spend time waiting
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent 20

Cloud Execution

Cloud providers handle resource management:

# Start conservative
harbor run --dataset terminal-bench@2.0 --agent claude-code -m anthropic/claude-opus-4-1 \
  --env daytona \
  --n-concurrent 50

# Scale up if no issues
harbor run --dataset terminal-bench@2.0 --agent claude-code -m anthropic/claude-opus-4-1 \
  --env daytona \
  --n-concurrent 200

Retry Configuration

Configure automatic retries for transient failures:

config.json

{
  "orchestrator": {
    "retry": {
      "max_retries": 3,
      "include_exceptions": null,  // Retry all exceptions
      "exclude_exceptions": [
        "AgentTimeoutError",      // Don't retry timeouts
        "VerifierTimeoutError",
        "RewardFileNotFoundError"
      ],
      "wait_multiplier": 2.0,      // Exponential backoff
      "min_wait_sec": 5.0,
      "max_wait_sec": 300.0
    }
  }
}

Retry Strategies

Aggressive (retry everything):

"retry": {
  "max_retries": 3,
  "exclude_exceptions": []
}

Conservative (retry only network/provider errors):

"retry": {
  "max_retries": 2,
  "include_exceptions": [
    "NetworkError",
    "ProviderError",
    "RateLimitError"
  ]
}

No retries:

"retry": {
  "max_retries": 0
}

Resource Management

Monitoring Resource Usage

# Monitor while running
watch -n 1 'docker stats'

# Check disk usage
df -h

# Monitor memory
free -h

Docker Resource Limits

Set global Docker resource limits:

~/.docker/daemon.json

{
  "default-ulimits": {
    "nofile": {
      "Name": "nofile",
      "Hard": 64000,
      "Soft": 64000
    }
  },
  "log-driver": "json-file",
  "log-opts": {
    "max-size": "10m",
    "max-file": "3"
  }
}

Restart Docker:

sudo systemctl restart docker

Disk Space Management

Large evaluations can consume significant disk space:

# Clean Docker images periodically
docker system prune -a -f

# Clean old job directories
rm -rf jobs/2026-02-*

# Monitor job directory size
du -sh jobs/

Task-Specific Resource Limits

Override default resources per task:

# Give tasks more resources
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --override-cpus 4 \
  --override-memory-mb 8192 \
  --override-storage-mb 20480

Overriding resources may disqualify results from benchmark leaderboards.

Optimizing Throughput

Pipeline Stages

Each trial goes through these stages:

Environment Build (5s - 10min)
Environment Start (2s - 30s)
Agent Setup (10s - 2min)
Agent Execution (1min - 1hr)
Verification (10s - 5min)
Cleanup (2s - 10s)

Reduce Build Time

Use pre-built images:

task.toml

[environment]
docker_image = "myregistry/task-env:v1.0"

Use Docker layer caching:

# Put slow, stable operations first
FROM python:3.11
RUN pip install numpy pandas  # Cached

# Put fast, changing operations last
COPY . /app

Use snapshots (cloud providers):

harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --env daytona \
  --ek snapshot_template_name="task-snapshot"

Reduce Startup Time

Minimize Dockerfile operations:

# Bad: Slow startup
FROM ubuntu:22.04
RUN apt-get update && apt-get install -y python3 python3-pip git
RUN pip install -r requirements.txt

# Good: Pre-built image
FROM myregistry/python-base:latest
COPY requirements.txt .
RUN pip install -r requirements.txt

Batch Similar Tasks

Group tasks with similar environments:

# Run all Python tasks together
harbor run --tasks ./tasks/python-* --agent claude-code -m anthropic/claude-opus-4-1

# Then run all JavaScript tasks
harbor run --tasks ./tasks/javascript-* --agent claude-code -m anthropic/claude-opus-4-1

Rate Limiting

LLM Provider Rate Limits

Stay within API rate limits:

# Anthropic: ~1000 RPM for Claude Opus
# With avg 10 requests per trial:
# 1000 RPM / 10 = 100 trials per minute
# 100 trials / 60 sec = ~1.5 concurrent trials

# Conservative: 25 concurrent
harbor run --dataset terminal-bench@2.0 --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent 25

# Aggressive: 100 concurrent (may hit limits)
harbor run --dataset terminal-bench@2.0 --agent claude-code -m anthropic/claude-opus-4-1 \
  --n-concurrent 100

Cloud Provider Rate Limits

# Daytona: Check your plan's concurrent sandbox limit
# Start conservative
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --env daytona \
  --n-concurrent 50

# Increase gradually
harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --env daytona \
  --n-concurrent 150

Progress Monitoring

Real-time Progress

Harbor shows real-time progress by default:

Running trials... ━━━━━━━━━━━━━━━━━━━ 45/100 45% 0:05:23

Active Trials (8/8):
  │ task-001__claude-code__1  →  Running agent...
  │ task-002__claude-code__1  →  Building environment...
  │ task-003__claude-code__1  →  Running verification...
  ...

Quiet Mode

For scripting or CI/CD:

harbor run --dataset my-tasks --agent claude-code -m anthropic/claude-opus-4-1 \
  --quiet

Logging

Check orchestrator logs:

tail -f jobs/2026-03-03__15-30-00/job.log

Debugging Parallel Execution

Common Issues

Too many concurrent trials:

Error: failed to create container: resource exhausted

Solution: Reduce --n-concurrent Rate limit errors:

Error: rate limit exceeded (429)

Solution: Reduce concurrency or add retry delays Out of memory:

Error: OOMKilled

Solution: Reduce concurrency or increase system memory Out of disk space:

Error: no space left on device

Solution: Clean up old jobs and Docker images

Debugging Individual Trials

When trials fail:

# View trial logs
harbor trials logs jobs/2026-03-03__15-30-00/trials/task-001__claude-code__1

# View trial results
harbor view jobs/2026-03-03__15-30-00/trials/task-001__claude-code__1

# Re-run single trial
harbor run --tasks ./tasks/task-001 \
  --agent claude-code \
  --model anthropic/claude-opus-4-1

Performance Benchmarks

Local Execution (16-core, 64GB RAM)

Concurrency	Tasks/Hour	CPU Usage	Memory Usage
4	~50	40%	16GB
8	~90	75%	32GB
16	~140	95%	56GB
32	~120	100%	60GB+

Diminishing returns beyond 16 concurrent trials

Cloud Execution (Daytona)

Concurrency	Tasks/Hour	Cost/Task
50	~600	$0.10
100	~1100	$0.09
200	~2000	$0.08

Approximate values, varies by task complexity

Best Practices

Start small: Test with 1-2 tasks before full evaluation
Increase gradually: Double concurrency until you hit limits
Monitor resources: Watch CPU, memory, and disk usage
Use appropriate environment: Local for fewer than 50 tasks, cloud for more than 100
Pre-build images: Use Docker layer caching and snapshots
Set appropriate timeouts: Avoid wasting resources on stuck trials
Enable retries: Handle transient failures automatically
Clean up regularly: Remove old jobs and Docker artifacts
Batch similar tasks: Group tasks with similar environments
Test retry config: Ensure retries work as expected

Example Configurations

Small Evaluation (Local)

{
  "orchestrator": {
    "type": "local",
    "n_concurrent_trials": 4,
    "retry": {
      "max_retries": 2
    }
  },
  "datasets": [
    {
      "registry": {"type": "remote"},
      "name": "terminal-bench",
      "version": "2.0",
      "n_tasks": 20
    }
  ]
}

Medium Evaluation (Local)

{
  "orchestrator": {
    "type": "local",
    "n_concurrent_trials": 12,
    "retry": {
      "max_retries": 3,
      "wait_multiplier": 2.0
    }
  },
  "timeout_multiplier": 1.5,
  "datasets": [
    {
      "registry": {"type": "remote"},
      "name": "terminal-bench",
      "version": "2.0"
    }
  ]
}

Large Evaluation (Cloud)

{
  "orchestrator": {
    "type": "local",
    "n_concurrent_trials": 100,
    "quiet": true,
    "retry": {
      "max_retries": 5,
      "wait_multiplier": 2.0,
      "max_wait_sec": 300.0
    }
  },
  "environment": {
    "type": "daytona",
    "kwargs": {
      "auto_delete_interval_mins": 10
    }
  },
  "datasets": [
    {
      "registry": {"type": "remote"},
      "name": "swebench",
      "version": "verified"
    }
  ]
}

Next Steps

Running Evaluations

Learn evaluation basics

Cloud Execution

Scale to cloud providers

Creating Tasks

Optimize tasks for parallel execution

​Execution Architecture

​Orchestrator Types

​Local Orchestrator

​Choosing Concurrency Level

​CPU-Bound Tasks

​Memory-Bound Tasks

​I/O-Bound Tasks

​Cloud Execution

​Retry Configuration

​Retry Strategies

​Resource Management

​Monitoring Resource Usage

​Docker Resource Limits

​Disk Space Management

​Task-Specific Resource Limits

​Optimizing Throughput

​Pipeline Stages

​Reduce Build Time

​Reduce Startup Time

​Batch Similar Tasks

​Rate Limiting

​LLM Provider Rate Limits

​Cloud Provider Rate Limits

​Progress Monitoring

​Real-time Progress

​Quiet Mode

​Logging

​Debugging Parallel Execution

​Common Issues

​Debugging Individual Trials

​Performance Benchmarks

​Local Execution (16-core, 64GB RAM)

​Cloud Execution (Daytona)

​Best Practices

​Example Configurations

​Small Evaluation (Local)

​Medium Evaluation (Local)

​Large Evaluation (Cloud)

​Next Steps

Running Evaluations

Cloud Execution

Creating Tasks

Execution Architecture

Orchestrator Types

Local Orchestrator

Choosing Concurrency Level

CPU-Bound Tasks

Memory-Bound Tasks

I/O-Bound Tasks

Cloud Execution

Retry Configuration

Retry Strategies

Resource Management

Monitoring Resource Usage

Docker Resource Limits

Disk Space Management

Task-Specific Resource Limits

Optimizing Throughput

Pipeline Stages

Reduce Build Time

Reduce Startup Time

Batch Similar Tasks

Rate Limiting

LLM Provider Rate Limits

Cloud Provider Rate Limits

Progress Monitoring

Real-time Progress

Quiet Mode

Logging

Debugging Parallel Execution

Common Issues

Debugging Individual Trials

Performance Benchmarks

Local Execution (16-core, 64GB RAM)

Cloud Execution (Daytona)

Best Practices

Example Configurations

Small Evaluation (Local)

Medium Evaluation (Local)

Large Evaluation (Cloud)

Next Steps