如何合理地估算线程池大小？

技术干货每日送达！

来源：蒋小强 ， ifeve.com/how-to-calculate-threadpool-size/

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));