如何合理地估算線程池大小？
先來一個天真的估算方法：
假設要求一個系統(tǒng)的TPS（Transaction Per Second或者Task Per Second）至少為20，然后假設每個Transaction由一個線程完成，繼續(xù)假設平均每個線程處理一個Transaction的時間為4s。那么問題轉化為：
如何設計線程池大小，使得可以在1s內處理完20個Transaction？
計算過程很簡單，每個線程的處理能力為0.25TPS，那么要達到20TPS，顯然需要20/0.25=80個線程。
然而這個方法有什么問題呢？
很顯然這個估算方法很天真，因為它沒有考慮到CPU數(shù)目。一般服務器的CPU核數(shù)為16或者32，如果有80個線程，那么肯定會帶來太多不必要的線程上下文切換開銷。
再來第二種簡單的但不知是否可行的方法（N為CPU總核數(shù)）：
如果是CPU密集型應用，則線程池大小設置為N+1
如果是IO密集型應用，則線程池大小設置為2N+1
如果一臺服務器上只部署這一個應用并且只有這一個線程池，那么這種估算或許合理，具體還需自行測試驗證。
接下來在這個文檔：服務器性能IO優(yōu)化 中發(fā)現(xiàn)一個估算公式：
最佳線程數(shù)目 = ((線程等待時間 + 線程CPU時間) / 線程CPU時間) * CPU數(shù)目
比如平均每個線程CPU運行時間為0.5s，而線程等待時間（非CPU運行時間，比如IO）為1.5s，CPU核心數(shù)為8，那么根據(jù)上面這個公式估算得到：((0.5+1.5)/0.5)*8=32。
這個公式進一步轉化為：
最佳線程數(shù)目 = (線程等待時間與線程CPU時間之比 + 1) * CPU數(shù)目
可以得出一個結論：
線程等待時間所占比例越高，需要越多線程。線程CPU時間所占比例越高，需要越少線程。
一個系統(tǒng)最快的部分是CPU，所以決定一個系統(tǒng)吞吐量上限的是CPU。增強CPU處理能力，可以提高系統(tǒng)吞吐量上限。但根據(jù)短板效應，真實的系統(tǒng)吞吐量并不能單純根據(jù)CPU來計算。那要提高系統(tǒng)吞吐量，就需要從“系統(tǒng)短板”（比如網(wǎng)絡延遲、IO）著手：
- 盡量提高短板操作的并行化比率，比如多線程下載技術
- 增強短板能力，比如用NIO替代IO
第一條可以聯(lián)系到Amdahl定律，這條定律定義了串行系統(tǒng)并行化后的加速比計算公式：
加速比 = 優(yōu)化前系統(tǒng)耗時 / 優(yōu)化后系統(tǒng)耗時
加速比越大，表明系統(tǒng)并行化的優(yōu)化效果越好。Addahl定律還給出了系統(tǒng)并行度、CPU數(shù)目和加速比的關系，加速比為Speedup，系統(tǒng)串行化比率（指串行執(zhí)行代碼所占比率）為F，CPU數(shù)目為N：
Speedup <= 1 / (F + (1 - F) / N)
當N足夠大時，串行化比率F越小，加速比Speedup越大。
那么，是否使用線程池就一定比使用單線程高效呢？
答案是否定的，比如Redis就是單線程的，但它卻非常高效，基本操作都能達到十萬量級/s。從線程這個角度來看，部分原因在于：
- 多線程帶來線程上下文切換開銷，單線程就沒有這種開銷
- 鎖
當然“Redis很快”更本質的原因在于：Redis基本都是內存操作，這種情況下單線程可以很高效地利用CPU。而多線程適用場景一般是：存在相當比例的IO和網(wǎng)絡操作。
所以即使有上面的簡單估算方法，也許看似合理，但實際上也未必合理，都需要結合系統(tǒng)真實情況（比如是IO密集型或者是CPU密集型或者是純內存操作）和硬件環(huán)境（CPU、內存、硬盤讀寫速度、網(wǎng)絡狀況等）來不斷嘗試達到一個符合實際的合理估算值。
最后來一個“Dark Magic”估算方法（因為我暫時還沒有搞懂它的原理），使用下面的類：
/**
* 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<Runnable> 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<Runnable> 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();
}
然后自己繼承這個抽象類并實現(xiàn)它的三個抽象方法，比如下面是我寫的一個示例（任務是請求網(wǎng)絡數(shù)據(jù)），其中我指定期望CPU利用率為1.0（即100%），任務隊列總大小不超過100,000字節(jié)：
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));
}
}
/**
* 自定義的異步IO任務
*
* @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();
}
}
}
得到的輸出如下：
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
推薦的任務隊列大小為2500，線程數(shù)為256，有點出乎意料之外。我可以如下構造一個線程池：
ThreadPoolExecutor pool = new ThreadPoolExecutor(6, 6, 0L, TimeUnit.MILLISECONDS,
new LinkedBlockingQueue<Runnable>(2500));
pool.setRejectedExecutionHandler(new ThreadPoolExecutor.CallerRunsPolicy());
參考：
