├── README.md ├── pom.xml └── src └── main └── java └── threadpool ├── PoolSizeCalculator.java └── SimplePoolSizeCaculator.java /README.md: -------------------------------------------------------------------------------- 1 | 2 | # 合理估算java的线程池大小及队列数 3 | ## 原理分析 4 | >原文：http://ifeve.com/how-to-calculate-threadpool-size/ 5 | 6 | 先来一个天真的估算方法：假设要求一个系统的TPS（Transaction Per Second或者Task Per Second）至少为20，然后假设每个Transaction由一个线程完成，继续假设平均每个线程处理一个Transaction的时间为4s。那么问题转化为： 7 | 8 | 如何设计线程池大小，使得可以在1s内处理完20个Transaction？ 9 | 10 | 计算过程很简单，每个线程的处理能力为0.25TPS，那么要达到20TPS，显然需要20/0.25=80个线程。 11 | 12 | 很显然这个估算方法很天真，因为它没有考虑到CPU数目。一般服务器的CPU核数为16或者32，如果有80个线程，那么肯定会带来太多不必要的线程上下文切换开销。 13 | 14 | 再来第二种简单的但不知是否可行的方法（N为CPU总核数）： 15 | 16 | - 如果是CPU密集型应用，则线程池大小设置为N+1； 17 | - 如果是IO密集型应用，则线程池大小设置为2N+1； 18 | 19 | 如果一台服务器上只部署这一个应用并且只有这一个线程池，那么这种估算或许合理，具体还需自行测试验证。 20 | 21 | 第三种方法是在服务器性能IO优化中发现的一个估算公式： 22 | 23 | >最佳线程数目 = （（线程等待时间 + 线程CPU时间）/ 线程CPU时间 ）* CPU数目 24 | 25 | 比如平均每个线程CPU运行时间为0.5s，而线程等待时间（非CPU运行时间，比如IO）为1.5s，CPU核心数为8，那么根据上面这个公式估算得到：((0.5+1.5)/0.5)*8=32。这个公式进一步转化为： 26 | 27 | >最佳线程数目 = （线程等待时间与线程CPU时间之比 + 1）* CPU数目 28 | 29 | 可以得出一个结论： 30 | 31 | - 线程CPU时间所占比例越高，需要越少线程。 32 | - 线程等待时间所占比例越高，需要越多线程。 33 | 34 | 上一种估算方法也和这个结论相合。 35 | 36 | 一个系统最快的部分是CPU，所以决定一个系统吞吐量上限的是CPU。增强CPU处理能力，可以提高系统吞吐量上限。但根据短板效应，真实的系统吞吐量并不能单纯根据CPU来计算。那要提高系统吞吐量，就需要从“系统短板”（比如网络延迟、IO）着手： 37 | 38 | - 尽量提高短板操作的并行化比率，比如多线程下载技术 39 | - 增强短板能力，比如用NIO替代IO 40 | 41 | 第一条可以联系到Amdahl定律，这条定律定义了串行系统并行化后的加速比计算公式： 42 | 43 | >加速比=优化前系统耗时 / 优化后系统耗时 44 | 45 | 加速比越大，表明系统并行化的优化效果越好。 46 | 47 | Addahl定律还给出了系统并行度、CPU数目和加速比的关系： 48 | 49 | 加速比为Speedup，系统串行化比率（指串行执行代码所占比率）为F，CPU数目为N： 50 | 51 | >Speedup <= 1 / (F + (1-F)/N) 52 | 53 | 当N足够大时，串行化比率F越小，加速比Speedup越大。 54 | 55 | 是否使用线程池就一定比使用单线程高效呢？ 56 | 57 | 答案是否定的，比如Redis就是单线程的，但它却非常高效，基本操作都能达到十万量级/s。 58 | 59 | 从线程这个角度来看，部分原因在于多线程带来线程上下文切换开销，单线程就没有这种开销。 60 | 61 | 当然“Redis很快”更本质的原因在于：Redis基本都是内存操作，这种情况下单线程可以很高效地利用CPU。而多线程适用场景一般是：存在相当比例的IO和网络操作。 62 | 63 | **所以即使有上面的估算方法，也许看似合理，但实际上也未必合理，都需要结合系统真实情况（比如是IO密集型或者是CPU密集型或者是纯内存操作）和硬件环境（CPU、内存、硬盘读写速度、网络状况等）来不断尝试达到一个符合实际的合理估算值。** 64 | 65 | ## 源码分析 66 | ### `PoolSizeCalculator`类 67 | #### `calculateBoundaries`方法 68 | 入口类，计算线程池大小和队列数。 69 | 70 | 接收两个参数，CPU负载和队列总内存的大小（bytes） 71 | #### `calculateMemoryUsage`方法 72 | 计算单个任务的内存大小，计算方法： 73 | 1. 手动GC 74 | 2. 计算可用内存大小m0 75 | 3. 创建一个队列，并往里面放1000个任务 76 | 4. 再次GC 77 | 5. 计算可用内存大小m1 78 | 6. `(m1 - m0) / 1000`即每个任务的大小 79 | >参考： 80 | #### `calculateOptimalCapacity`方法 81 | 计算队列数，计算公式：**队列总内存**/**单个任务的内存**。 82 | 83 | 接收一个参数：队列总内存的大小。 84 | 85 | #### `start`方法 86 | 计算*执行3秒的任务*所消耗CPU的实际使用时间。 87 | >参考： 88 | #### `calculateOptimalThreadCount`方法 89 | 计算线程池大小。 90 | 计算公式：CPU核数 * （1 + 线程等待时间/线程CPU时间） 91 | #### `collectGarbage`方法 92 | 循环手动GC 93 | ### `SimplePoolSizeCaculator`类 94 | `PoolSizeCalculator`类的一个实现，计算*CPU负载1*，*队列总内存的大小为100k左右*的*IO密集型*的线程池大小和队列数 95 | #### `AsyncIOTask`类 96 | IO密集型的一个例子 97 | 98 | ## 使用方法 99 | ```bash 100 | # 下载 101 | git clone https://github.com/sunshanpeng/dark_magic.git 102 | # 编译 103 | cd dark_magic && mvn package 104 | # 执行 105 | java -jar target/dark_magic-1.0-SNAPSHOT.jar 106 | ``` 107 | 控制台打印示例 108 | ```$xslt 109 | Target queue memory usage (bytes): 100000 110 | createTask() produced threadpool.AsyncIOTask which took 40 bytes in a queue 111 | Formula: 100000 / 40 112 | * Recommended queue capacity (bytes): 2500 113 | Number of CPU: 4 114 | Target utilization: 1 115 | Elapsed time (nanos): 3000000000 116 | Compute time (nanos): 125000000 117 | Wait time (nanos): 2875000000 118 | Formula: 4 * 1 * (1 + 2875000000 / 125000000) 119 | * Optimal thread count: 96 120 | ``` 121 | > 如果不修改队列内存大小和任务，队列数可能都是2500 -------------------------------------------------------------------------------- /pom.xml: -------------------------------------------------------------------------------- 1 | 3 | 4.0.0 4 | 5 | threadpool 6 | dark_magic 7 | 1.0-SNAPSHOT 8 | jar 9 | dark_magic 10 | 11 | 12 | UTF-8 13 | 14 | 15 | 16 | 17 | 18 | 19 | org.apache.maven.plugins 20 | maven-jar-plugin 21 | 3.1.0 22 | 23 | 24 | 25 | true 26 | lib/ 27 | threadpool.SimplePoolSizeCaculator 28 | 29 | 30 | 31 | 32 | 33 | org.apache.maven.plugins 34 | maven-compiler-plugin 35 | 36 | 6 37 | 6 38 | 39 | 40 | 41 | 42 | 43 | 44 | -------------------------------------------------------------------------------- /src/main/java/threadpool/PoolSizeCalculator.java: -------------------------------------------------------------------------------- 1 | package threadpool; 2 | 3 | import java.math.BigDecimal; 4 | import java.math.RoundingMode; 5 | import java.util.Timer; 6 | import java.util.TimerTask; 7 | import java.util.concurrent.BlockingQueue; 8 | 9 | /** 10 | * A class that calculates the optimal thread pool boundaries. It takes the 11 | * desired target utilization and the desired work queue memory consumption as 12 | * input and retuns thread count and work queue capacity. 13 | * 14 | * @author Niklas Schlimm 15 | */ 16 | public abstract class PoolSizeCalculator { 17 | 18 | /** 19 | * The sample queue size to calculate the size of a single {@link Runnable} 20 | * element. 21 | */ 22 | private static final int SAMPLE_QUEUE_SIZE = 1000; 23 | 24 | /** 25 | * Accuracy of test run. It must finish within 20ms of the testTime 26 | * otherwise we retry the test. This could be configurable. 27 | */ 28 | private static final int EPSYLON = 20; 29 | 30 | /** 31 | * Control variable for the CPU time investigation. 32 | */ 33 | private volatile boolean expired; 34 | 35 | /** 36 | * Time (millis) of the test run in the CPU time calculation. 37 | */ 38 | private final long elapsed = 3000; 39 | 40 | /** 41 | * Calculates the boundaries of a thread pool for a given {@link Runnable}. 42 | * 43 | * @param targetUtilization the desired utilization of the CPUs (0 <= targetUtilization <= 1) 44 | * @param targetQueueSizeBytes the desired maximum work queue size of the thread pool (bytes) 45 | */ 46 | void calculateBoundaries(BigDecimal targetUtilization, BigDecimal targetQueueSizeBytes) { 47 | calculateOptimalCapacity(targetQueueSizeBytes); 48 | Runnable task = createTask(); 49 | start(task); 50 | start(task); // warm up phase 51 | long cputime = getCurrentThreadCPUTime(); 52 | start(task); // test interval 53 | cputime = getCurrentThreadCPUTime() - cputime; 54 | long waitTime = (elapsed * 1000000) - cputime; 55 | calculateOptimalThreadCount(cputime, waitTime, targetUtilization); 56 | } 57 | 58 | private void calculateOptimalCapacity(BigDecimal targetQueueSizeBytes) { 59 | long mem = calculateMemoryUsage(); 60 | BigDecimal queueCapacity = targetQueueSizeBytes.divide(new BigDecimal(mem), 61 | RoundingMode.HALF_UP); 62 | System.out.println("Target queue memory usage (bytes): " 63 | + targetQueueSizeBytes); 64 | System.out.println("createTask() produced " + createTask().getClass().getName() + " which took " + mem + " bytes in a queue"); 65 | System.out.println("Formula: " + targetQueueSizeBytes + " / " + mem); 66 | System.out.println("* Recommended queue capacity (bytes): " + queueCapacity); 67 | } 68 | 69 | /** 70 | * Brian Goetz' optimal thread count formula, see 'Java Concurrency in 71 | * * Practice' (chapter 8.2) * 72 | * * @param cpu 73 | * * cpu time consumed by considered task 74 | * * @param wait 75 | * * wait time of considered task 76 | * * @param targetUtilization 77 | * * target utilization of the system 78 | */ 79 | private void calculateOptimalThreadCount(long cpu, long wait, 80 | BigDecimal targetUtilization) { 81 | BigDecimal computeTime = new BigDecimal(cpu); 82 | BigDecimal waitTime = new BigDecimal(wait); 83 | BigDecimal numberOfCPU = new BigDecimal(Runtime.getRuntime() 84 | .availableProcessors()); 85 | BigDecimal optimalthreadcount = numberOfCPU.multiply(targetUtilization) 86 | .multiply(new BigDecimal(1).add(waitTime.divide(computeTime, 87 | RoundingMode.HALF_UP))); 88 | System.out.println("Number of CPU: " + numberOfCPU); 89 | System.out.println("Target utilization: " + targetUtilization); 90 | System.out.println("Elapsed time (nanos): " + (elapsed * 1000000)); 91 | System.out.println("Compute time (nanos): " + cpu); 92 | System.out.println("Wait time (nanos): " + wait); 93 | System.out.println("Formula: " + numberOfCPU + " * " 94 | + targetUtilization + " * (1 + " + waitTime + " / " 95 | + computeTime + ")"); 96 | System.out.println("* Optimal thread count: " + optimalthreadcount); 97 | } 98 | 99 | /** 100 | * * Runs the {@link Runnable} over a period defined in {@link #elapsed}. 101 | * * Based on Heinz Kabbutz' ideas 102 | * * (http://www.javaspecialists.eu/archive/Issue124.html). 103 | * * 104 | * * @param task 105 | * * the runnable under investigation 106 | */ 107 | public void start(Runnable task) { 108 | long start = 0; 109 | int runs = 0; 110 | do { 111 | if (++runs > 10) { 112 | throw new IllegalStateException("Test not accurate"); 113 | } 114 | expired = false; 115 | start = System.currentTimeMillis(); 116 | Timer timer = new Timer(); 117 | timer.schedule(new TimerTask() { 118 | public void run() { 119 | expired = true; 120 | } 121 | }, elapsed); 122 | while (!expired) { 123 | task.run(); 124 | } 125 | start = System.currentTimeMillis() - start; 126 | timer.cancel(); 127 | } while (Math.abs(start - elapsed) > EPSYLON); 128 | collectGarbage(3); 129 | } 130 | 131 | private void collectGarbage(int times) { 132 | for (int i = 0; i < times; i++) { 133 | System.gc(); 134 | try { 135 | Thread.sleep(10); 136 | } catch (InterruptedException e) { 137 | Thread.currentThread().interrupt(); 138 | break; 139 | } 140 | } 141 | } 142 | 143 | /** 144 | * Calculates the memory usage of a single element in a work queue. Based on 145 | * Heinz Kabbutz' ideas 146 | * (http://www.javaspecialists.eu/archive/Issue029.html). 147 | * 148 | * @return memory usage of a single {@link Runnable} element in the thread 149 | * pools work queue 150 | */ 151 | private long calculateMemoryUsage() { 152 | BlockingQueue queue = createWorkQueue(SAMPLE_QUEUE_SIZE); 153 | for (int i = 0; i < SAMPLE_QUEUE_SIZE; i++) { 154 | queue.add(createTask()); 155 | } 156 | long mem0 = Runtime.getRuntime().totalMemory() - 157 | Runtime.getRuntime().freeMemory(); 158 | long mem1 = Runtime.getRuntime().totalMemory() - 159 | Runtime.getRuntime().freeMemory(); 160 | queue = null; 161 | collectGarbage(15); 162 | mem0 = Runtime.getRuntime().totalMemory() 163 | - Runtime.getRuntime().freeMemory(); 164 | queue = createWorkQueue(SAMPLE_QUEUE_SIZE); 165 | for (int i = 0; i < SAMPLE_QUEUE_SIZE; i++) { 166 | queue.add(createTask()); 167 | } 168 | collectGarbage(15); 169 | mem1 = Runtime.getRuntime().totalMemory() 170 | - Runtime.getRuntime().freeMemory(); 171 | return (mem1 - mem0) / SAMPLE_QUEUE_SIZE; 172 | } 173 | 174 | /** 175 | * Create your runnable task here. 176 | * 177 | * @return an instance of your runnable task under investigation 178 | */ 179 | protected abstract Runnable createTask(); 180 | 181 | /** 182 | * Return an instance of the queue used in the thread pool. 183 | * 184 | * @return queue instance 185 | */ 186 | protected abstract BlockingQueue createWorkQueue(int capacity); 187 | 188 | /** 189 | * Calculate current cpu time. Various frameworks may be used here, 190 | * depending on the operating system in use. (e.g. 191 | * http://www.hyperic.com/products/sigar). The more accurate the CPU time 192 | * measurement, the more accurate the results for thread count boundaries. 193 | * 194 | * @return current cpu time of current thread 195 | */ 196 | protected abstract long getCurrentThreadCPUTime(); 197 | 198 | } -------------------------------------------------------------------------------- /src/main/java/threadpool/SimplePoolSizeCaculator.java: -------------------------------------------------------------------------------- 1 | package threadpool; 2 | 3 | import java.io.BufferedReader; 4 | import java.io.IOException; 5 | import java.io.InputStreamReader; 6 | import java.lang.management.ManagementFactory; 7 | import java.math.BigDecimal; 8 | import java.net.HttpURLConnection; 9 | import java.net.URL; 10 | import java.util.concurrent.BlockingQueue; 11 | import java.util.concurrent.LinkedBlockingQueue; 12 | 13 | public class SimplePoolSizeCaculator extends PoolSizeCalculator { 14 | 15 | @Override 16 | protected Runnable createTask() { 17 | return new AsyncIOTask(); 18 | } 19 | 20 | @Override 21 | protected BlockingQueue createWorkQueue(int capacity) { 22 | return new LinkedBlockingQueue(capacity); 23 | } 24 | 25 | @Override 26 | protected long getCurrentThreadCPUTime() { 27 | //the total CPU time for the current thread in nanoseconds 28 | return ManagementFactory.getThreadMXBean().getCurrentThreadCpuTime(); 29 | } 30 | 31 | public static void main(String[] args) { 32 | PoolSizeCalculator poolSizeCalculator = new SimplePoolSizeCaculator(); 33 | poolSizeCalculator.calculateBoundaries(new BigDecimal(1.0), new BigDecimal(100000)); 34 | } 35 | 36 | } 37 | 38 | /** 39 | * 自定义的异步IO任务 40 | * @author Will 41 | * 42 | */ 43 | class AsyncIOTask implements Runnable { 44 | 45 | @Override 46 | public void run() { 47 | HttpURLConnection connection = null; 48 | BufferedReader reader = null; 49 | try { 50 | URL url = new URL("http://baidu.com"); 51 | 52 | connection = (HttpURLConnection) url.openConnection(); 53 | connection.connect(); 54 | reader = new BufferedReader(new InputStreamReader( 55 | connection.getInputStream())); 56 | 57 | String line; 58 | StringBuilder stringBuilder; 59 | while ((line = reader.readLine()) != null) { 60 | stringBuilder = new StringBuilder(); 61 | stringBuilder.append(line); 62 | } 63 | } 64 | 65 | catch (IOException e) { 66 | 67 | } finally { 68 | if(reader != null) { 69 | try { 70 | reader.close(); 71 | } 72 | catch(Exception e) { 73 | 74 | } 75 | } 76 | if (connection != null) 77 | connection.disconnect(); 78 | } 79 | 80 | } 81 | 82 | } --------------------------------------------------------------------------------