Memcached 集群，客户端自动Hash到不同服务器的实现

.....
首先分析一下Java client 启动时的部分代码

Memcached 支持直接设置多个servers属性 来实现多个memcahced均衡，对应还有一个属性是weights，字面意思就是权重，分析了一下代码，和我想的是一样的

启动memcached的代码通常是这样的

SockIOPool pool = SockIOPool.getInstance(poolname);
pool.setServers(servers);
pool.setWeights(weights);
pool.setInitConn(initConn);
pool.setMinConn(minConn);
pool.setMaxConn(maxConn);
pool.setMaxIdle(maxIdle);
pool.setMaxBusyTime(maxBusyTime);
pool.setMaintSleep(maintSleep);
pool.setSocketTO(socketTO);
pool.setSocketConnectTO(socketConnectTO);
pool.setNagle(nagle);
pool.setHashingAlg(SockIOPool.NEW_COMPAT_HASH);
pool.initialize();
MemCachedClient client = new MemCachedClient(poolname);

servers 和 weights 都是一个数组，就是说可以同时设置多个server

然后看一下 pool.initialize() 做了什么

availPool   = new HashMap<String,Map<SockIO,Long>>( servers.length * initConn );
busyPool    = new HashMap<String,Map<SockIO,Long>>( servers.length * initConn );
deadPool    = new IdentityHashMap<SockIO,Integer>();
hostDeadDur = new HashMap<String,Long>();
hostDead    = new HashMap<String,Date>();
maxCreate   = (poolMultiplier > minConn) ? minConn : minConn / poolMultiplier;	// only create up to maxCreate connections at once
if ( log.isDebugEnabled() ) {
	log.debug( "++++ initializing pool with following settings:" );
	log.debug( "++++ initial size: " + initConn );
	log.debug( "++++ min spare   : " + minConn );
	log.debug( "++++ max spare   : " + maxConn );
}
// if servers is not set, or it empty, then
// throw a runtime exception
if ( servers == null || servers.length <= 0 ) {
	log.error( "++++ trying to initialize with no servers" );
	throw new IllegalStateException( "++++ trying to initialize with no servers" );
}
	// initalize our internal hashing structures
        if ( this.hashingAlg == CONSISTENT_HASH )
		populateConsistentBuckets();
	else
	        populateBuckets();

看到这里就是开辟一些连接池的空间，然后调用了根据我们选择的hash 算法 执行populateBuckets();或者populateConsistentBuckets();

hash算法共有4种

// native String.hashCode();
public static final int NATIVE_HASH     = 0;	
// original compatibility hashing algorithm (works with other clients)
public static final int OLD_COMPAT_HASH = 1;
// new CRC32 based compatibility hashing algorithm (works with other clients)	
public static final int NEW_COMPAT_HASH = 2;
// MD5 Based -- Stops thrashing when a server added or removed
public static final int CONSISTENT_HASH = 3;		

我们通常用的是 NEW_COMPAT_HASH，这个保证可以wokrs with other clients

所以看一下populateBuckets()做了什么

this.buckets = new ArrayList<String>();
for ( int i = 0; i < servers.length; i++ ) {
  if ( this.weights != null && this.weights.length > i ) {
     for ( int k = 0; k < this.weights[i].intValue(); k++ ) {
	 this.buckets.add( servers[i] );
	 if ( log.isDebugEnabled() )
	    log.debug( "++++ added " + servers[i] + " to server bucket" );
     }
  }
  else {
     this.buckets.add( servers[i] );
  }
     // create initial connections
  for ( int j = 0; j < initConn; j++ ) {
          SockIO socket = createSocket( servers[i] );
	  if ( socket == null ) {
		break;
          }
          addSocketToPool( availPool, servers[i], socket );
  }
}

假如我们设置的servers是 192.168.0.1:44444和192.168.0.2:22222
然后我们设置了weights是 5和3 那么
buckets list的值最终会是
[
192.168.0.1:44444,
192.168.0.1:44444,
192.168.0.1:44444,
192.168.0.1:44444,
192.168.0.1:44444,
192.168.0.2:22222,
192.168.0.2:22222.
192.168.0.2:22222.
]
然后就开始根据initCon初始连接数按servers分别创建socket
那么究竟这个buckets做什么用呢？

在我们使用set存放对象时会调用

SockIOPool.SockIO sock = pool.getSock( key, hashCode );

看一看pool.getSock的代码

// get initial bucket
long bucket = getBucket( key, hashCode );
String server = ( this.hashingAlg == CONSISTENT_HASH )
		? consistentBuckets.get( bucket )
		: buckets.get( (int)bucket );

其中有段代码是这样的，看看getBucket

private long getBucket( String key, Integer hashCode ) {
	long hc = getHash( key, hashCode );
	if ( this.hashingAlg == CONSISTENT_HASH ) {
		return findPointFor( hc );
	}
	else {
	     long bucket = hc % buckets.size();
		if ( bucket < 0 ) bucket *= -1;
			return bucket;
	}
}

先不管key和hashCode，我们看到首先算出一个hc值后会直接做hc%buckets.size()实际上就是根据buckets的数量散列，最终值一定是buckets.size()范围里的一个值
然后最终server值就根据buckets.get( (int)bucket )得到，那么假如我们得到bucket是3，则参照上面buckets 里的值，得到 list.get(3)=192.168.0.1:44444 所以会根据weight设置的值的不同得到不同的server ，如果 weights设置10:1 那buckets里就是10个相同的server和另一个不同的，将来散列得到的server很大可能性是servers里设置的第一个server。