0x6 Java系列：如何合理地估算线程池大小？

如何合理地估算线程池大小？

这个问题虽然看起来很小，却并不那么容易回答。大家如果有更好的方法欢迎赐教，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：

如何设计线程池大小，使得可以在1s内处理完20个Transaction？

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。

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。

再来第二种简单的但不知是否可行的方法（N为CPU总核数）：

·      如果是CPU密集型应用，则线程池大小设置为N+1

·      如果是IO密集型应用，则线程池大小设置为2N+1

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。

接下来在这个文档：服务器性能IO优化中发现一个估算公式：

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

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：

最佳线程数目= （线程等待时间与线程CPU时间之比+ 1）* CPU数目

可以得出一个结论：

线程等待时间所占比例越高，需要越多线程。线程CPU时间所占比例越高，需要越少线程。

上一种估算方法也和这个结论相合。

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：

·      尽量提高短板操作的并行化比率，比如多线程下载技术

·      增强短板能力，比如用NIO替代IO

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：

加速比=优化前系统耗时/ 优化后系统耗时

加速比越大，表明系统并行化的优化效果越好。Addahl定律还给出了系统并行度、CPU数目和加速比的关系，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：

Speedup <= 1 / (F + (1-F)/N)

当N足够大时，串行化比率F越小，加速比Speedup越大。

写到这里，我突然冒出一个问题。

是否使用线程池就一定比使用单线程高效呢？

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。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

·      锁

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。

所以即使有上面的简单估算方法，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。

最后来一个“Dark Magic”估算方法（因为我暂时还没有搞懂它的原理），使用下面的类：

package pool_size_calculate;

import java.math.BigDecimal;

import java.math.RoundingMode;

import java.util.Timer;

import java.util.TimerTask;

import java.util.concurrent.BlockingQueue;

/**

 * A class that calculates the optimal thread pool boundaries. It takes the

 * desired target utilization and the desired work queue memory consumption as

 * input and retuns thread count and work queue capacity.

 *

 * @author Niklas Schlimm

 *

 */

public abstract class PoolSizeCalculator {

       /**

        * The sample queue size to calculate the size of a single {@link Runnable}

        * element.

        */

       private final int SAMPLE_QUEUE_SIZE = 1000;

       /**

        * Accuracy of test run. It must finish within 20ms of the testTime

        * otherwise we retry the test. This could be configurable.

        */

       private final int EPSYLON = 20;

       /**

        * Control variable for the CPU time investigation.

        */

       private volatile boolean expired;

       /**

        * Time (millis) of the test run in the CPU time calculation.

        */

       private final long testtime = 3000;

       /**

        * Calculates the boundaries of a thread pool for a given {@link Runnable}.

        *

        * @param targetUtilization

        *            the desired utilization of the CPUs (0 <= targetUtilization <=       *            1)        * @param targetQueueSizeBytes         *            the desired maximum work queue size of the thread pool (bytes)  */    protected void calculateBoundaries(BigDecimal targetUtilization,                 BigDecimal targetQueueSizeBytes) {             calculateOptimalCapacity(targetQueueSizeBytes);           Runnable task = creatTask();              start(task);            start(task); // warm up phase             long cputime = getCurrentThreadCPUTime();            start(task); // test intervall           cputime = getCurrentThreadCPUTime() - cputime;               long waittime = (testtime * 1000000) - cputime;        calculateOptimalThreadCount(cputime, waittime, targetUtilization);       }      private void calculateOptimalCapacity(BigDecimal targetQueueSizeBytes) {             long mem = calculateMemoryUsage();               BigDecimal queueCapacity = targetQueueSizeBytes.divide(new BigDecimal(                       mem), RoundingMode.HALF_UP);              System.out.println("Target queue memory usage (bytes): "                             + targetQueueSizeBytes);            System.out.println("createTask() produced "                          + creatTask().getClass().getName() + " which took " + mem                        + " bytes in a queue");        System.out.println("Formula: " + targetQueueSizeBytes + " / " + mem);        System.out.println("* Recommended queue capacity (bytes): "                            + queueCapacity); }      /**     * Brian Goetz' optimal thread count formula, see 'Java Concurrency in      * Practice' (chapter 8.2)       *      * @param cpu        *            cpu time consumed by considered task         * @param wait         *            wait time of considered task    * @param targetUtilization    *            target utilization of the system        */    private void calculateOptimalThreadCount(long cpu, long wait,                 BigDecimal targetUtilization) {            BigDecimal waitTime = new BigDecimal(wait);          BigDecimal computeTime = new BigDecimal(cpu);           BigDecimal numberOfCPU = new BigDecimal(Runtime.getRuntime()                           .availableProcessors());               BigDecimal optimalthreadcount = numberOfCPU.multiply(targetUtilization)                        .multiply(                                           new BigDecimal(1).add(waitTime.divide(computeTime,                                                      RoundingMode.HALF_UP)));        System.out.println("Number of CPU: " + numberOfCPU);        System.out.println("Target utilization: " + targetUtilization);            System.out.println("Elapsed time (nanos): " + (testtime * 1000000));              System.out.println("Compute time (nanos): " + cpu);               System.out.println("Wait time (nanos): " + wait);              System.out.println("Formula: " + numberOfCPU + " * "                          + targetUtilization + " * (1 + " + waitTime + " / "                         + computeTime + ")");          System.out.println("* Optimal thread count: " + optimalthreadcount);     }      /**     * Runs the {@link Runnable} over a period defined in {@link #testtime}.     * Based on Heinz Kabbutz' ideas        * (http://www.javaspecialists.eu/archive/Issue124.html).  *          * @param task       *            the runnable under investigation     */    public void start(Runnable task) {              long start = 0;              int runs = 0;          do {               if (++runs > 5) {

                            throw new IllegalStateException("Test not accurate");

                     }

                     expired = false;

                     start = System.currentTimeMillis();

                     Timer timer = new Timer();

                     timer.schedule(new TimerTask() {

                            public void run() {

                                   expired = true;

                            }

                     }, testtime);

                     while (!expired) {

                            task.run();

                     }

                     start = System.currentTimeMillis() - start;

                     timer.cancel();

              } while (Math.abs(start - testtime) > EPSYLON);

              collectGarbage(3);

       }

       private void collectGarbage(int times) {

              for (int i = 0; i < times; i++) {

                     System.gc();

                     try {

                            Thread.sleep(10);

                     } catch (InterruptedException e) {

                            Thread.currentThread().interrupt();

                            break;

                     }

              }

       }

       /**

        * Calculates the memory usage of a single element in a work queue. Based on

        * Heinz Kabbutz' ideas

        * (http://www.javaspecialists.eu/archive/Issue029.html).

        *

        * @return memory usage of a single {@link Runnable} element in the thread

        *         pools work queue

        */

       public long calculateMemoryUsage() {

              BlockingQueue queue = createWorkQueue();

              for (int i = 0; i < SAMPLE_QUEUE_SIZE; i++) {

                     queue.add(creatTask());

              }

              long mem0 = Runtime.getRuntime().totalMemory()

                            - Runtime.getRuntime().freeMemory();

              long mem1 = Runtime.getRuntime().totalMemory()

                            - Runtime.getRuntime().freeMemory();

              queue = null;

              collectGarbage(15);

              mem0 = Runtime.getRuntime().totalMemory()

                            - Runtime.getRuntime().freeMemory();

              queue = createWorkQueue();

              for (int i = 0; i < SAMPLE_QUEUE_SIZE; i++) {

                     queue.add(creatTask());

              }

              collectGarbage(15);

              mem1 = Runtime.getRuntime().totalMemory()

                            - Runtime.getRuntime().freeMemory();

              return (mem1 - mem0) / SAMPLE_QUEUE_SIZE;

       }

       /**

        * Create your runnable task here.

        *

        * @return an instance of your runnable task under investigation

        */

       protected abstract Runnable creatTask();

       /**

        * Return an instance of the queue used in the thread pool.

        *

        * @return queue instance

        */

       protected abstract BlockingQueue createWorkQueue();

       /**

        * Calculate current cpu time. Various frameworks may be used here,

        * depending on the operating system in use. (e.g.

        * http://www.hyperic.com/products/sigar). The more accurate the CPU time

        * measurement, the more accurate the results for thread count boundaries.

        *

        * @return current cpu time of current thread

        */

       protected abstract long getCurrentThreadCPUTime();

}

然后自己继承这个抽象类并实现它的三个抽象方法，比如下面是我写的一个示例（任务是请求网络数据），其中我指定期望CPU利用率为1.0（即100%），任务队列总大小不超过100,000字节：

package pool_size_calculate;

import java.io.BufferedReader;

import java.io.IOException;

import java.io.InputStreamReader;

import java.lang.management.ManagementFactory;

import java.math.BigDecimal;

import java.net.HttpURLConnection;

import java.net.URL;

import java.util.concurrent.BlockingQueue;

import java.util.concurrent.LinkedBlockingQueue;

public class SimplePoolSizeCaculatorImpl extends PoolSizeCalculator {

       @Override

       protected Runnable creatTask() {

              return new AsyncIOTask();

       }

       @Override

       protected BlockingQueue createWorkQueue() {

              return new LinkedBlockingQueue(1000);

       }

       @Override

       protected long getCurrentThreadCPUTime() {

              return ManagementFactory.getThreadMXBean().getCurrentThreadCpuTime();

       }

       public static void main(String[] args) {

              PoolSizeCalculator poolSizeCalculator = new SimplePoolSizeCaculatorImpl();

              poolSizeCalculator.calculateBoundaries(new BigDecimal(1.0), new BigDecimal(100000));

       }

}

/**

 * 自定义的异步IO任务

 * @author Will

 *

 */

class AsyncIOTask implements Runnable {

       @Override

       public void run() {

              HttpURLConnection connection = null;

              BufferedReader reader = null;

              try {

                     String getURL = "http://baidu.com";

                     URL getUrl = new URL(getURL);

                     connection = (HttpURLConnection) getUrl.openConnection();

                     connection.connect();

                     reader = new BufferedReader(new InputStreamReader(

                                   connection.getInputStream()));

                     String line;

                     while ((line = reader.readLine()) != null) {

                            // empty loop

                     }

              }

              catch (IOException e) {

              } finally {

                     if(reader != null) {

                            try {

                                   reader.close();

                            }

                            catch(Exception e) {

                            }

                     }

                     connection.disconnect();

              }

       }

}

得到的输出如下：

Target queue memory usage (bytes): 100000

createTask() produced pool_size_calculate.AsyncIOTask which took 40 bytes in a queue

Formula: 100000 / 40

* Recommended queue capacity (bytes): 2500

Number of CPU: 4

Target utilization: 1

Elapsed time (nanos): 3000000000

Compute time (nanos): 47181000

Wait time (nanos): 2952819000

Formula: 4 * 1 * (1 + 2952819000 / 47181000)

* Optimal thread count: 256

推荐的任务队列大小为2500，线程数为256，有点出乎意料之外。我可以如下构造一个线程池：

ThreadPoolExecutor pool =

 new ThreadPoolExecutor(256, 256, 0L, TimeUnit.MILLISECONDS, new LinkedBlockingQueue(2500));

文章来源：蒋小强 ，

ifeve.com/how-to-calculate-threadpool-size/