JVM没有提供这样的API，即使对于以开头的代理也是如此
-javaagent。JVM
TI是为以该
-agent选项开头的本机代理或调试器提供的本机接口。Java代理可能会使用Instrumentation API，该API提供了类检测的底层功能，但没有直接分析功能。
有两种类型的性能分析实现，即通过采样和通过仪器。
采样通过定期记录堆栈跟踪信息（样本）来进行。这不会跟踪每个方法调用，但仍会检测到热点，因为热点在记录的堆栈跟踪中多次出现。优点是它不需要代理程序或特殊的API，并且您可以控制探查器的开销。您可以通过ThreadMXBean来实现它，该工具允许您获取所有正在运行的线程的堆栈跟踪。实际上，即使a
Thread.getAllStackTraces()也可以，但是会
ThreadMXBean提供有关线程的更详细的信息。
因此，主要任务是为堆栈跟踪中找到的方法实现有效的存储结构，即将同一方法的出现折叠为单个调用树项。
这是一个非常简单的采样器在自己的JVM上工作的示例：
import java.lang.Thread.State;import java.lang.management.ManagementFactory;import java.lang.management.ThreadInfo;import java.lang.management.ThreadMXBean;import java.util.*;import java.util.concurrent.Executors;import java.util.concurrent.ScheduledExecutorService;import java.util.concurrent.TimeUnit;public class Sampler { private static final ThreadMXBean TMX=ManagementFactory.getThreadMXBean(); private static String CLASS, METHOD; private static CallTree ROOT; private static ScheduledExecutorService EXECUTOR; public static synchronized void startSampling(String className, String method) { if(EXECUTOR!=null) throw new IllegalStateException("sampling in progress"); System.out.println("sampling started"); CLASS=className; METHOD=method; EXECUTOR = Executors.newScheduledThreadPool(1); // "fixed delay" reduces overhead, "fixed rate" raises precision EXECUTOR.scheduleWithFixedDelay(new Runnable() { public void run() { newSample(); } }, 150, 75, TimeUnit.MILLISECONDS); } public static synchronized CallTree stopSampling() throws InterruptedException { if(EXECUTOR==null) throw new IllegalStateException("no sampling in progress"); EXECUTOR.shutdown(); EXECUTOR.awaitTermination(Long.MAX_VALUE, TimeUnit.DAYS); EXECUTOR=null; final CallTree root = ROOT; ROOT=null; return root; } public static void printCallTree(CallTree t) { if(t==null) System.out.println("method not seen"); else printCallTree(t, 0, 100); } private static void printCallTree(CallTree t, int ind, long percent) { long num=0; for(CallTree ch:t.values()) num+=ch.count; if(num==0) return; for(Map.Entry<List<String>,CallTree> ch:t.entrySet()) { CallTree cht=ch.getValue(); StringBuilder sb = new StringBuilder(); for(int p=0; p<ind; p++) sb.append(' '); final long chPercent = cht.count*percent/num; sb.append(chPercent).append("% (").append(cht.cpu*percent/num) .append("% cpu) ").append(ch.getKey()).append(" "); System.out.println(sb.toString()); printCallTree(cht, ind+2, chPercent); } } static class CallTree extends HashMap<List<String>, CallTree> { long count=1, cpu; CallTree(boolean cpu) { if(cpu) this.cpu++; } CallTree getOrAdd(String cl, String m, boolean cpu) { List<String> key=Arrays.asList(cl, m); CallTree t=get(key); if(t!=null) { t.count++; if(cpu) t.cpu++; } else put(key, t=new CallTree(cpu)); return t; } } static void newSample() { for(ThreadInfo ti:TMX.dumpAllThreads(false, false)) { final boolean cpu = ti.getThreadState()==State.RUNNABLE; StackTraceElement[] stack=ti.getStackTrace(); for(int ix = stack.length-1; ix>=0; ix--) { StackTraceElement ste = stack[ix]; if(!ste.getClassName().equals(CLASS)||!ste.getMethodName().equals(METHOD)) continue; CallTree t=ROOT; if(t==null) ROOT=t=new CallTree(cpu); for(ix--; ix>=0; ix--) { ste = stack[ix]; t=t.getOrAdd(ste.getClassName(), ste.getMethodName(), cpu); } } } }}探查器在不通过调试API的情况下搜寻每个方法调用，使用工具将通知代码添加到他们感兴趣的每个方法中。优点是它们永远不会错过方法调用，但另一方面，它们却在执行中增加了大量开销搜索热点时可能会影响结果。而且实施起来更加复杂。我无法为您提供此类字节代码转换的代码示例。
Instrumentation
API仅提供给Java代理，但是如果您想进入Instrumentation方向，下面的程序演示了如何连接到其自己的JVM并作为Java代理加载自身：
import java.io.*;import java.lang.instrument.Instrumentation;import java.lang.management.ManagementFactory;import java.nio.ByteBuffer;import java.nio.charset.Charset;import java.nio.charset.StandardCharsets;import java.util.UUID;import java.util.zip.ZipEntry;import java.util.zip.ZipOutputStream;// this API comes from the tools.jar of your JDKimport com.sun.tools.attach.*;public class SelfAttacher { public static Instrumentation BACK_link; public static void main(String[] args) throws Exception { // create a special property to verify our JVM connection String magic=UUID.randomUUID().toString()+'/'+System.nanoTime(); System.setProperty("magic", magic); // the easiest way uses the non-standardized runtime name string String name=ManagementFactory.getRuntimeMXBean().getName(); int ix=name.indexOf('@'); if(ix>=0) name=name.substring(0, ix); VirtualMachine vm; getVM: { try { vm = VirtualMachine.attach(name); if(magic.equals(vm.getSystemProperties().getProperty("magic"))) break getVM; } catch(Exception ex){} // if the easy way failed, try iterating over all local JVMs for(VirtualMachineDescriptor vd:VirtualMachine.list()) try { vm=VirtualMachine.attach(vd); if(magic.equals(vm.getSystemProperties().getProperty("magic"))) break getVM; vm.detach(); } catch(Exception ex){} // could not find our own JVM or could not attach to it return; } System.out.println("attached to: "+vm.id()+'/'+vm.provider().type()); vm.loadAgent(createJar().getAbsolutePath()); synchronized(SelfAttacher.class) { while(BACK_link==null) SelfAttacher.class.wait(); } System.out.println("Now I have hands on instrumentation: "+BACK_link); System.out.println(BACK_link.isModifiableClass(SelfAttacher.class)); vm.detach(); } // create a JAR file for the agent; since our class is already in class path // our jar consisting of a MANIFEST declaring our class as agent only private static File createJar() throws IOException { File f=File.createTempFile("agent", ".jar"); f.deleteonExit(); Charset cs=StandardCharsets.ISO_8859_1; try(FileOutputStream fos=new FileOutputStream(f); ZipOutputStream os=new ZipOutputStream(fos)) { os.putNextEntry(new ZipEntry("meta-INF/MANIFEST.MF")); ByteBuffer bb = cs.enpre("Agent-Class: "+SelfAttacher.class.getName()); os.write(bb.array(), bb.arrayOffset()+bb.position(), bb.remaining()); os.write(10); os.closeEntry(); } return f; } // invoked when the agent is loaded into the JVM, pass inst back to the caller public static void agentmain(String agentArgs, Instrumentation inst) { synchronized(SelfAttacher.class) { BACK_link=inst; SelfAttacher.class.notifyAll(); } }}
