parallel-processing-patterns

关键指南

Windows 文件路径要求

强制要求：在 Windows 上的文件路径中始终使用反斜杠

在 Windows 上使用 Edit 或 Write 工具时，文件路径必须使用反斜杠（\），而不是正斜杠（/）。

---

Bash 并行处理模式（2025）

概述

本指南全面介绍 Bash 中的并行与并发执行，涵盖 GNU Parallel、xargs 并行化、作业控制、工作池以及现代异步模式，助你实现最高性能。

GNU Parallel（推荐）

安装

# Debian/Ubuntu
sudo apt-get install parallel

# macOS
brew install parallel

# From source
wget https://ftp.gnu.org/gnu/parallel/parallel-latest.tar.bz2
tar -xjf parallel-latest.tar.bz2
cd parallel-*
./configure && make && sudo make install

基本用法

#!/usr/bin/env bash
set -euo pipefail

# Process multiple files in parallel
parallel gzip ::: *.txt

# Equivalent to:
# for f in *.txt; do gzip "$f"; done
# But runs in parallel!

# Using find with parallel
find . -name "*.jpg" | parallel convert {} -resize 50% resized/{}

# Specify number of jobs
parallel -j 8 process_file ::: *.dat

# From stdin
cat urls.txt | parallel -j 10 wget -q

# Multiple inputs
parallel echo ::: A B C ::: 1 2 3
# Output: A 1, A 2, A 3, B 1, B 2, B 3, C 1, C 2, C 3

# Paired inputs with :::+
parallel echo ::: A B C :::+ 1 2 3
# Output: A 1, B 2, C 3

输入处理

#!/usr/bin/env bash
set -euo pipefail

# Input from file
parallel -a input.txt process_line

# Multiple input files
parallel -a file1.txt -a file2.txt 'echo {1} {2}'

# Column-based input
cat data.tsv | parallel --colsep '\t' 'echo Name: {1}, Value: {2}'

# Named columns
cat data.csv | parallel --header : --colsep ',' 'echo {name}: {value}'

# Null-delimited for safety with special characters
find . -name "*.txt" -print0 | parallel -0 wc -l

# Line-based chunking
cat huge_file.txt | parallel --pipe -N1000 'wc -l'

替换字符串

#!/usr/bin/env bash
set -euo pipefail

# {} - Full input
parallel echo 'Processing: {}' ::: file1.txt file2.txt

# {.} - Remove extension
parallel echo '{.}' ::: file.txt file.csv
# Output: file, file

# {/} - Basename
parallel echo '{/}' ::: /path/to/file.txt
# Output: file.txt

# {//} - Directory path
parallel echo '{//}' ::: /path/to/file.txt
# Output: /path/to

# {/.} - Basename without extension
parallel echo '{/.}' ::: /path/to/file.txt
# Output: file

# {#} - Job number (1-based)
parallel echo 'Job {#}: {}' ::: A B C

# {%} - Slot number (recycled job slot)
parallel -j 2 'echo "Slot {%}: {}"' ::: A B C D E

# Combined
parallel 'convert {} -resize 50% {//}/thumb_{/.}.jpg' ::: *.png

进度与日志记录

#!/usr/bin/env bash
set -euo pipefail

# Show progress bar
parallel --bar process_item ::: {1..100}

# Progress with ETA
parallel --progress process_item ::: {1..100}

# Verbose output
parallel --verbose gzip ::: *.txt

# Log to file
parallel --joblog jobs.log gzip ::: *.txt

# Resume from where it left off (skip completed jobs)
parallel --joblog jobs.log --resume gzip ::: *.txt

# Results logging
parallel --results results_dir 'echo {1} + {2}' ::: 1 2 3 ::: 4 5 6
# Creates: results_dir/1/4/stdout, results_dir/1/4/stderr, etc.

资源管理

#!/usr/bin/env bash
set -euo pipefail

# CPU-based parallelism (number of cores)
parallel -j "$(nproc)" process_item ::: {1..1000}

# Leave some cores free
parallel -j '-2' process_item ::: {1..1000}  # nproc - 2

# Percentage of cores
parallel -j '50%' process_item ::: {1..1000}

# Load-based throttling
parallel --load 80% process_item ::: {1..1000}

# Memory-based throttling
parallel --memfree 2G process_item ::: {1..1000}

# Rate limiting (max jobs per second)
parallel -j 4 --delay 0.5 wget ::: url1 url2 url3 url4

# Timeout per job
parallel --timeout 60 long_process ::: {1..100}

# Retry failed jobs
parallel --retries 3 flaky_process ::: {1..100}

分布式执行

#!/usr/bin/env bash
set -euo pipefail

# Run on multiple servers
parallel --sshloginfile servers.txt process_item ::: {1..1000}

# servers.txt format:
# 4/server1.example.com  (4 jobs on server1)
# 8/server2.example.com  (8 jobs on server2)
# :                       (local machine)

# Transfer files before execution
parallel --sshloginfile servers.txt --transferfile {} process {} ::: *.dat

# Return results
parallel --sshloginfile servers.txt --return {.}.result process {} ::: *.dat

# Cleanup after transfer
parallel --sshloginfile servers.txt --transfer --return {.}.out --cleanup \
    'process {} > {.}.out' ::: *.dat

# Environment variables
export MY_VAR="value"
parallel --env MY_VAR --sshloginfile servers.txt 'echo $MY_VAR' ::: A B C

复杂管道

#!/usr/bin/env bash
set -euo pipefail

# Pipe mode - distribute stdin across workers
cat huge_file.txt | parallel --pipe -N1000 'sort | uniq -c'

# Block size for pipe mode
cat data.bin | parallel --pipe --block 10M 'process_chunk'

# Keep order of output
parallel --keep-order 'sleep $((RANDOM % 3)); echo {}' ::: A B C D E

# Group output (don't mix output from different jobs)
parallel --group 'for i in 1 2 3; do echo "Job {}: line $i"; done' ::: A B C

# Tag output with job identifier
parallel --tag 'echo "output from {}"' ::: A B C

# Sequence output (output as they complete, but grouped)
parallel --ungroup 'echo "Starting {}"; sleep 1; echo "Done {}"' ::: A B C

xargs 并行化

xargs 基本并行用法

#!/usr/bin/env bash
set -euo pipefail

# -P for parallel jobs
find . -name "*.txt" | xargs -P 4 -I {} gzip {}

# -n for items per command
echo {1..100} | xargs -n 10 -P 4 echo "Batch:"

# Null-delimited for safety
find . -name "*.txt" -print0 | xargs -0 -P 4 -I {} process {}

# Multiple arguments per process
cat urls.txt | xargs -P 10 -n 5 wget -q

# Limit max total arguments
echo {1..1000} | xargs -P 4 --max-args=50 echo

使用 xargs 执行复杂命令

#!/usr/bin/env bash
set -euo pipefail

# Use sh -c for complex commands
find . -name "*.jpg" -print0 | \
    xargs -0 -P 4 -I {} sh -c 'convert "$1" -resize 50% "thumb_$(basename "$1")"' _ {}

# Multiple placeholders
paste file1.txt file2.txt | \
    xargs -P 4 -n 2 sh -c 'diff "$1" "$2" > "diff_$(basename "$1" .txt).patch"' _

# Process in batches
find . -name "*.log" -print0 | \
    xargs -0 -P 4 -n 100 tar -czvf logs_batch.tar.gz

# With failure handling
find . -name "*.dat" -print0 | \
    xargs -0 -P 4 -I {} sh -c 'process "$1" || echo "Failed: $1" >> failures.log' _ {}

作业控制模式

后台作业管理

#!/usr/bin/env bash
set -euo pipefail

# Track background jobs
declare -a PIDS=()

# Start jobs
for item in {1..10}; do
    process_item "$item" &
    PIDS+=($!)
done

# Wait for all
for pid in "${PIDS[@]}"; do
    wait "$pid"
done

echo "All jobs complete"

# Or wait for any to complete
wait -n  # Bash 4.3+
echo "At least one job complete"

使用信号量管理作业池

#!/usr/bin/env bash
set -euo pipefail

# Maximum concurrent jobs
MAX_JOBS=4

# Simple semaphore using a counter
job_count=0

run_with_limit() {
    local cmd=("$@")

    # Wait if at limit
    while ((job_count >= MAX_JOBS)); do
        wait -n 2>/dev/null || true
        ((job_count--))
    done

    # Start new job
    "${cmd[@]}" &
    ((job_count++))
}

# Usage
for item in {1..20}; do
    run_with_limit process_item "$item"
done

# Wait for remaining
wait

基于 FIFO 的作业池

#!/usr/bin/env bash
set -euo pipefail

MAX_JOBS=4
JOB_FIFO="/tmp/job_pool_$$"

# Create job slots
mkfifo "$JOB_FIFO"
trap 'rm -f "$JOB_FIFO"' EXIT

# Initialize slots
exec 3<>"$JOB_FIFO"
for ((i=0; i<MAX_JOBS; i++)); do
    echo >&3
done

# Run with slot
run_with_slot() {
    local cmd=("$@")

    read -u 3  # Acquire slot (blocks if none available)

    {
        "${cmd[@]}"
        echo >&3  # Release slot
    } &
}

# Usage
for item in {1..20}; do
    run_with_slot process_item "$item"
done

wait
exec 3>&-

工作池模式

#!/usr/bin/env bash
set -euo pipefail

WORK_QUEUE="/tmp/work_queue_$$"
RESULT_QUEUE="/tmp/result_queue_$$"
NUM_WORKERS=4

mkfifo "$WORK_QUEUE" "$RESULT_QUEUE"
trap 'rm -f "$WORK_QUEUE" "$RESULT_QUEUE"' EXIT

# Worker function
worker() {
    local id="$1"
    while read -r task; do
        [[ "$task" == "STOP" ]] && break

        # Process task
        local result
        result=$(process_task "$task" 2>&1)
        echo "RESULT:$id:$task:$result"
    done
}

# Start workers
for ((i=0; i<NUM_WORKERS; i++)); do
    worker "$i" < "$WORK_QUEUE" > "$RESULT_QUEUE" &
done

# Result collector (background)
collect_results() {
    while read -r line; do
        [[ "$line" == "DONE" ]] && break
        echo "$line" >> results.txt
    done < "$RESULT_QUEUE"
} &
COLLECTOR_PID=$!

# Producer - send work
{
    for task in "${TASKS[@]}"; do
        echo "$task"
    done

    # Stop signals for workers
    for ((i=0; i<NUM_WORKERS; i++)); do
        echo "STOP"
    done
} > "$WORK_QUEUE"

# Signal end of results
wait  # Wait for workers
echo "DONE" > "$RESULT_QUEUE"
wait "$COLLECTOR_PID"

现代异步模式与性能优化

关于 PID 收集、wait -n、有界并发、快速失败的作业池、进度报告、信号安全清理、批次大小、CPU 数量调优、内存感知并发和基准测试的更详细示例，参见 references/modern-async-patterns.md 和 references/performance-optimization.md。对于超出 GNU Parallel / xargs / 作业控制核心模式的复杂编排或调优，请加载这些参考资料。

错误处理

优雅的失败处理

#!/usr/bin/env bash
set -euo pipefail

# Track failures
declare -A FAILURES

parallel_with_retry() {
    local max_retries=3
    local items=("$@")

    for item in "${items[@]}"; do
        local retries=0
        local success=false

        while ((retries < max_retries)) && ! $success; do
            if process_item "$item"; then
                success=true
            else
                ((retries++))
                echo "Retry $retries for $item" >&2
                sleep $((retries * 2))  # Exponential backoff
            fi
        done

        if ! $success; then
            FAILURES["$item"]="Failed after $max_retries retries"
        fi
    done &

    wait
}

# Report failures
report_failures() {
    if ((${#FAILURES[@]} > 0)); then
        echo "Failures:" >&2
        for item in "${!FAILURES[@]}"; do
            echo "  $item: ${FAILURES[$item]}" >&2
        done
        return 1
    fi
}

取消支持

#!/usr/bin/env bash
set -euo pipefail

# Global cancellation flag
CANCELLED=false
declare -a WORKER_PIDS=()

cancel_all() {
    CANCELLED=true
    for pid in "${WORKER_PIDS[@]}"; do
        kill "$pid" 2>/dev/null || true
    done
}

trap cancel_all SIGINT SIGTERM

cancellable_worker() {
    local id="$1"
    while ! $CANCELLED; do
        # Check for work
        if work=$(get_next_work); then
            process_work "$work"
        else
            sleep 0.1
        fi
    done
}

# Start workers
for ((i=0; i<NUM_WORKERS; i++)); do
    cancellable_worker "$i" &
    WORKER_PIDS+=($!)
done

# Wait with interrupt support
wait || true

资源

• GNU Parallel 教程
• GNU Parallel 手册
• Bash 作业控制
• 高级 Bash 脚本指南 - 进程替换

---

掌握并行处理，高效利用多核并加快脚本执行速度。