dask-optimization（DASK 优化）

Dask - Advanced Optimization & Cluster Tuning

Parallel computing is not "free". In a distributed environment, the cost of moving data (network I/O) and scheduling tasks can often exceed the computation time. This guide focuses on minimizing overhead and maximizing throughput.

When to Use

• Your Dask jobs are failing with "KilledWorker" or "OutOfMemory" errors.
• The Dask Dashboard shows a lot of "red" (communication) or "gray" (idle) time.
• You need to process datasets that are 10x-100x larger than the total RAM of your cluster.
• You are building custom distributed algorithms using dask.delayed or Futures.
• You need to optimize resource allocation (CPU vs. Threads) for specific workloads.

Reference Documentation

• Best Practices: https://docs.dask.org/en/latest/best-practices.html
• Distributed Diagnostics: https://distributed.dask.org/en/latest/diagnosing-performance.html
• Memory Management: https://distributed.dask.org/en/latest/worker.html#memory-management
• Search patterns: client.scatter, dask.compute(optimize_graph=True), repartition, client.restart

Core Principles

1. Communication is the Killer

The fastest distributed task is the one that doesn't need data from another machine. Aim for data locality.

2. The Goldilocks Chunk Size

• Too small: Scheduler is overwhelmed by millions of tiny tasks (Task Overhead).
• Too large: Tasks don't fit in memory, causing disk spilling or worker crashes.
• Target: 100MB - 300MB per chunk for most numeric data.

3. Computation vs. Serialization

Every object sent to a worker must be serialized (pickled). Large Python objects (like complex dicts) passed as arguments can slow down the cluster significantly.

Quick Reference: Performance Profiling

from dask.distributed import Client, performance_report

client = Client("tcp://scheduler-address:8786")

# Generate a detailed HTML report of the computation
with performance_report(filename="dask-report.html"):
    result = big_computation.compute()

# Tip: Check the "Task Stream" for gaps. Gaps mean workers are idle
# waiting for the scheduler or network.

Critical Rules

✅ DO

• Use client.scatter - If multiple tasks need the same large piece of data, send it to workers once.
• Prefer map_partitions - In DataFrames, this allows you to run a single optimized pandas operation per chunk instead of row-wise logic.
• Use persist() for branching - If you use the same intermediate result in two different computations, persist() it in memory to avoid re-calculating the entire graph twice.
• Profile with the Dashboard - Watch the "Memory" and "Task Stream" tabs in real-time.
• Match Threads to Workload - Use many threads for I/O bound tasks (web scraping, reading files) and 1 thread per worker for CPU-bound tasks (NumPy math, ML training) to avoid Global Interpreter Lock (GIL) issues.

❌ DON'T

• Don't use compute() in loops - This pulls data to the local machine and destroys parallel efficiency.
• Don't pass Large Data as Arguments - Instead of delayed(func)(large_df), use large_future = client.scatter(large_df) then delayed(func)(large_future).
• Don't ignore Data Skew - If one worker has 10GB of data and others have 100MB, the cluster is only as fast as the slowest worker. Use repartition or rechunk.
• Don't use list(dask_collection) - This forces an immediate compute of all elements into local memory.

Data Locality & Communication

Using scatter and Futures

# ❌ BAD: Large object sent to every task (High Overhead)
large_lookup = load_heavy_dict()
results = [delayed(process)(x, large_lookup) for x in data]

# ✅ GOOD: Scatter once, use reference
[large_future] = client.scatter([large_lookup], broadcast=True)
results = [delayed(process)(x, large_future) for x in data]

Memory Management Tuning

Fighting the "KilledWorker"

Worker memory has thresholds:

• Target (0.6): Dask tries to stay below this.
• Spill (0.7): Dask starts moving data to disk.
• Pause (0.8): Worker stops accepting new tasks.
• Terminate (0.95): OS or Dask kills the worker.

# Adjusting worker limits via config (distributed.yaml or code)
import dask
dask.config.set({"distributed.worker.memory.target": 0.45})
dask.config.set({"distributed.worker.memory.spill": 0.55})

Task Graph Optimization

Fusing Operations

Dask automatically "fuses" many operations into one task to reduce scheduler overhead.

# Multiple operations on a DataArray
da = da + 1
da = da * 2
da = da.sum()

# When computing, Dask optimizes the graph
# You can inspect it:
da.visualize(filename='graph.pdf', optimize_graph=True)

Practical Workflows

1. Optimizing Large Joins (Shuffling)

Joins are expensive because they require moving data between workers (shuffling).

def optimized_join(left_ddf, right_ddf):
    # 1. Ensure right side is small enough to be broadcasted
    # or ensure both are partitioned by the join key.

    # If one DF is small (e.g. < 100MB)
    # result = left_ddf.map_partitions(lambda df: df.merge(small_df_local, on='key'))

    # 2. If both are large, set the index first (triggers a shuffle once)
    left_ddf = left_ddf.set_index('key')
    right_ddf = right_ddf.set_index('key')

    # 3. Subsequent joins will be "locally aligned" (Zero communication)
    return left_ddf.merge(right_ddf, left_index=True, right_index=True)

2. High-Throughput I/O with Parquet

def fast_save(ddf, path):
    # 1. Categorical columns save massive space
    ddf = ddf.categorize()

    # 2. Write with efficient compression
    # 'snappy' is usually the best balance for speed
    ddf.to_parquet(path, engine='pyarrow', compression='snappy',
                   write_metadata_file=True)

3. Managing a Long-running Cluster

# Prevent a "dirty" cluster from slowing down
def clean_cluster_state():
    client.cancel(list(client.futures)) # Clear all references
    client.restart() # Hard reset all workers
    import gc
    gc.collect() # Local cleanup

Advanced Configuration

Resource Tagging

Tell Dask to run specific tasks only on specific workers (e.g., those with a GPU).

# Start worker with: dask-worker ... --resources "GPU=1"

# Submit task requesting resource
future = client.submit(my_gpu_function, data, resources={'GPU': 1})

Common Pitfalls and Solutions

The "Zombie Worker" (serialization error)

If your task requires a library that isn't installed on the workers, the task will fail repeatedly.

# ✅ Solution: Use pip_install or conda_install via client
from dask.distributed import PipInstall
client.register_plugin(PipInstall(packages=["scikit-learn"]))

Unmanaged Memory

Python's garbage collector isn't immediate. Sometimes workers appear full because "Unmanaged Memory" hasn't been freed.

# ✅ Solution: Manually trigger GC on workers
def worker_gc():
    import gc
    return gc.collect()

client.run(worker_gc)

Too Many Partitions

If you have 10,000 partitions for a 1GB dataset, you'll spend more time scheduling than calculating.

# ✅ Solution: Repartition to fewer pieces
ddf = ddf.repartition(npartitions=20)

Dask Optimization is the art of balancing resources. By understanding the flow of data through the network and the mechanics of worker memory, you can scale Python logic to planetary-scale datasets with industrial reliability.