代理服务第一版速度挺猛，但是满载的时候内存消耗到无法忍受，最高的时候吃了我1.8G内存，后经过一系列调优操作，目前满载内存消耗约为100MB。
首先，Netty本身的零拷贝技术理论上内存消耗不大，但是经过对jvm堆分析发现netty的ByteBuf占用了大量的空间，查阅文档后发现Netty默认会开启缓存策略，将对象复用来提高性能，我的需求是性能差不多得了，内存消耗低才是王道，所以关了。
其次，我分析jvm堆的时候还发现占用最高的是byte[]对象，而大量使用byte对象的场景只有编解码以及加解密的时候，虽然byte[]用完就丢gc也能处理，但是众所周知gc并不会及时清理垃圾，而且频繁gc代价可不小。
所以我使用了对象池技术，byte[]对象只能在对象池中获取，对象池大小可以在极小和极大值之间动态扩张。
代码语言：javascript复制 protected void encode(ChannelHandlerContext ctx, TcpMessage msg, ByteBuf out) throws Exception {
if (msg.getBodyBuf() != null) {
byte[] data=msg.getBodyBuf();
byte[] ciphertext=ByteArrayPool.getIns().getByteArray(data.length);
EncryptUtil.AESEncode(data, data.length, ciphertext, configProperties.getDesKey());
msg.setEncryptBodyBuf(ciphertext);
ByteArrayPool.getIns().returnByteArray(data);
}
byte[] headBuf = msg.getfHead().Serialize();
byte[] ciphertext=ByteArrayPool.getIns().getByteArray(headBuf.length);
EncryptUtil.AESEncode(headBuf, headBuf.length, ciphertext, configProperties.getDesKey());
EncryptUtil.selfWriteByte(out,ciphertext,ciphertext.length,configProperties);
if (msg.getBodyBuf() != null) {
EncryptUtil.selfWriteByte(out,msg.getBodyBuf(),msg.getBodyBuf().length,configProperties);
}
ByteArrayPool.getIns().returnByteArray(headBuf);
ByteArrayPool.getIns().returnByteArray(msg.getBodyBuf());
}
protected void decode(ChannelHandlerContext ctx, ByteBuf in, List<Object> out) throws Exception {
TcpMessage tcpMessage = new TcpMessage();
try {
in.markReaderIndex();
if (in.readableBytes() < TcpHead.headLength) {
in.resetReaderIndex();
return;
}
TcpHead tcpHead = new TcpHead();
byte[] ciphertext = ByteArrayPool.getIns().getByteArray(TcpHead.headLength);
byte[] headByte = ByteArrayPool.getIns().getByteArray(TcpHead.headLength, true);
if (!EncryptUtil.selfReadByte(in,ciphertext,ciphertext.length,configProperties)){
//读取失败，长度不够
in.resetReaderIndex();
ByteArrayPool.getIns().returnByteArray(ciphertext);
ByteArrayPool.getIns().returnByteArray(headByte);
return;
}
EncryptUtil.AESDecode(ciphertext, ciphertext.length, headByte, configProperties.getDesKey());
tcpHead.DeSerialize(ByteBuffer.wrap(headByte));
ByteArrayPool.getIns().returnByteArray(ciphertext);
ByteArrayPool.getIns().returnByteArray(headByte);
tcpMessage.setfHead(tcpHead);
if (tcpHead.getCrc() != tcpHead.calCrc()) {
ctx.channel().close();
System.out.println("CRC校验失败");
}
if (in.readableBytes() < tcpHead.getLength()) {
in.resetReaderIndex();
return;
}
if (tcpHead.getRealLength() > 0) {
byte[] mainbuf = ByteArrayPool.getIns().getByteArray(tcpHead.getRealLength());
ciphertext = ByteArrayPool.getIns().getByteArray(tcpHead.getLength(), true);
if (!EncryptUtil.selfReadByte(in,ciphertext,ciphertext.length,configProperties)){
//读取失败，长度不够
in.resetReaderIndex();
ByteArrayPool.getIns().returnByteArray(ciphertext);
ByteArrayPool.getIns().returnByteArray(mainbuf);
return;
}
EncryptUtil.AESDecode(ciphertext, ciphertext.length, mainbuf, configProperties.getDesKey());
ByteArrayPool.getIns().returnByteArray(ciphertext);
tcpMessage.setBodyBuf(mainbuf, tcpHead.getRealLength());
}
out.add(tcpMessage);
} catch (Exception ex) {
// Logger.LogError(ex.getMessage(), ex);
}
}public class ByteArrayPool {
int minN = 100;
int maxN = 500;
int mod = 16;
private static class Ins {
private static ByteArrayPool I = new ByteArrayPool();
}
public static ByteArrayPool getIns() {
return Ins.I;
}
private Map<Integer, LinkedBlockingQueue<byte[]>> pool = new ConcurrentHashMap<>();
public byte[] getByteArray(int length, boolean isNotUpper) {
if (isNotUpper) return get(length);
length = mod - length % mod;
return get(length);
}
public byte[] getByteArray(int length) {
return getByteArray(length, false);
}
private byte[] get(int length) {
byte[] ret;
poolDataMap.computeIfAbsent(length,k->new AtomicInteger(0));
pool.computeIfAbsent(length, k -> new LinkedBlockingQueue<>());
LinkedBlockingQueue<byte[]> queue = pool.get(length);
try {
if (poolDataMap.get(length).get()<=minN){
poolDataMap.get(length).addAndGet(1);
return new byte[length];
}
ret= queue.poll();
if (ret==null){
if (length<50){
poolDataMap.get(length).addAndGet(1);
return new byte[length];
}
Thread.sleep(10);
poolDataMap.get(length).addAndGet(1);
return new byte[length];
}
return ret;
} catch (Throwable exception) {
poolDataMap.get(length).addAndGet(1);
return new byte[length];
}
}
public static Map<Integer, AtomicInteger> poolDataMap=new ConcurrentHashMap<>();
public void returnByteArray(byte[] bytes) {
if (bytes == null) return;
pool.computeIfAbsent(bytes.length, k -> new LinkedBlockingQueue<>());
LinkedBlockingQueue<byte[]> queue = pool.get(bytes.length);
if (queue.size() > maxN) {
poolDataMap.computeIfAbsent(bytes.length,k->new AtomicInteger(0));
poolDataMap.get(bytes.length).addAndGet(-1);
System.gc();
return;
}
queue.offer(bytes);
}
}另外cipher.doFinal是可以传入byte[]来接收加解密结果的，这样我们加解密时候cipher就不会浪费我们多余空间了。
代码语言：javascript复制 cipher.doFinal(data, 0, length, result, 0);最后就是jvm启动的时候，我换了cms来做gc，因为cms提供了一个fullGc的参数，老年代内存占用达到临界值就强行fullGc，而且cms管理小内存貌似不错，我把临界值设置为70%。
做完这些后把jvm堆大小上限设置为100M，测试后发现我的500M带宽能跑满，内存占用不到100M，收工！
