摘要: 在stackoverflow上有一个非常有名的问题:为什么处理有序数组要比非有序数组快?,可见分支预测对代码运行效率有非常大的影响。要提高代码执行效率,一个重要的原则就是尽量避免CPU把流水线清空,那么提高分支预测的成功率就非常重要。
分支预测
在stackoverflow上有一个非常有名的问题:为什么处理有序数组要比非有序数组快?,可见分支预测对代码运行效率有非常大的影响。
现代CPU都支持分支预测(branch prediction)和指令流水线(instruction pipeline),这两个结合可以极大提高CPU效率。对于像简单的if跳转,CPU是可以比较好地做分支预测的。但是对于switch跳转,CPU则没有太多的办法。switch本质上是据索引,从地址数组里取地址再跳转。
要提高代码执行效率,一个重要的原则就是尽量避免CPU把流水线清空,那么提高分支预测的成功率就非常重要。
那么对于代码里,如果某个switch分支概率很高,是否可以考虑代码层面帮CPU把判断提前,来提高代码执行效率呢?
Dubbo里ChannelEventRunnable的switch判断
在ChannelEventRunnable
里有一个switch来判断channel state,然后做对应的逻辑:查看
一个channel建立起来之后,超过99.9%情况它的state都是ChannelState.RECEIVED
,那么可以考虑把这个判断提前。
benchmark验证
下面通过jmh来验证下:
public class TestBenchMarks {
public enum ChannelState { CONNECTED, DISCONNECTED, SENT, RECEIVED, CAUGHT } @State(Scope.Benchmark) public static class ExecutionPlan { @Param({ "1000000" }) public int size; public ChannelState[] states = null; @Setup public void setUp() { ChannelState[] values = ChannelState.values(); states = new ChannelState[size]; Random random = new Random(new Date().getTime()); for (int i = 0; i < size; i++) { int nextInt = random.nextInt(1000000); if (nextInt > 100) { states[i] = ChannelState.RECEIVED; } else { states[i] = values[nextInt % values.length]; } } } } @Fork(value = 5) @Benchmark @BenchmarkMode(Mode.Throughput) public void benchSiwtch(ExecutionPlan plan, Blackhole bh) { int result = 0; for (int i = 0; i < plan.size; ++i) { switch (plan.states[i]) { case CONNECTED: result += ChannelState.CONNECTED.ordinal(); break; case DISCONNECTED: result += ChannelState.DISCONNECTED.ordinal(); break; case SENT: result += ChannelState.SENT.ordinal(); break; case RECEIVED: result += ChannelState.RECEIVED.ordinal(); break; case CAUGHT: result += ChannelState.CAUGHT.ordinal(); break; } } bh.consume(result); } @Fork(value = 5) @Benchmark @BenchmarkMode(Mode.Throughput) public void benchIfAndSwitch(ExecutionPlan plan, Blackhole bh) { int result = 0; for (int i = 0; i < plan.size; ++i) { ChannelState state = plan.states[i]; if (state == ChannelState.RECEIVED) { result += ChannelState.RECEIVED.ordinal(); } else { switch (state) { case CONNECTED: result += ChannelState.CONNECTED.ordinal(); break; case SENT: result += ChannelState.SENT.ordinal(); break; case DISCONNECTED: result += ChannelState.DISCONNECTED.ordinal(); break; case CAUGHT: result += ChannelState.CAUGHT.ordinal(); break; } } } bh.consume(result); }
}
- benchSiwtch里是纯switch判断
- benchIfAndSwitch 里用一个if提前判断state是否
ChannelState.RECEIVED
benchmark结果是:
Result "io.github.hengyunabc.jmh.TestBenchMarks.benchSiwtch": 576.745 ±(99.9%) 6.806 ops/s [Average] (min, avg, max) = (490.348, 576.745, 618.360), stdev = 20.066 CI (99.9%): 569.939, 583.550
Run complete. Total time: 00:06:48
Benchmark (size) Mode Cnt Score Error Units
TestBenchMarks.benchIfAndSwitch 1000000 thrpt 100 1535.867 ± 61.212 ops/s
TestBenchMarks.benchSiwtch 1000000 thrpt 100 576.745 ± 6.806 ops/s
可以看到提前if判断的确提高了代码效率,这种技巧可以放在性能要求严格的地方。
Benchmark代码:https://github.com/hengyunabc/jmh-demo
总结
- switch对于CPU来说难以做分支预测
- 某些switch条件如果概率比较高,可以考虑单独提前if判断,充分利用CPU的分支预测机制
原文地址:http://blog.51cto.com/13679539/2128966