Memcached(简称为:MC)在互联网广泛使用,是最基础的架构。但MC的mget(即一次获取多个值)一直是一个难题,我们的要求是mget性能上要尽量接近普通memcache get。下面通过一段伪代码介绍了如何以接近get single value的性能实现mget,并且就该架构在实际环境中遇到的一些问题加以讨论。
场景
在开始这个话题之前先考虑一个问题,为什么需要MC mget?Redis不是已经很好的实现了list,hashset,hashtable,zset等等丰富的数据结构吗?这个问题需要从本厂的应用场景开始。用户登陆之后会修改自己的状态,同时获得自己关注人的状态。修改自己的状态是一次MC set过程。自己的关注人列表可以从Redis中获得,此时key是用户的uid,value是关注任的list。获得自己关注人的状态则是根据关注人uid的一次MC get,时间复杂度是O(1)。可以这样做,在程序中执行一个for循环,依次从MC中get关注人状态,这个get过程的时间复杂度是O(n)。当关注人列表扩展到2000时,每次MC get平均耗时2~5ms,这种线性循环获取好友状态的办法要耗时10s,是完全无法接受的。怎么解决这个问题呢?
通过NIO实现mget,并发的执行MC get
danga.memcached2.0.1已经使用NIO框架来实现mget,但是它的实现有些问题,参考:http://blog.youkuaiyun.com/e_wsq/article/details/7876801。mget伪代码如下:
private final class Conn {
public ByteBuffer outgoing;
// 使用一个ByteBuffer list来存储从MC读出的内容
public List<ByteBuffer> incoming = new ArrayList<ByteBuffer>();
public Conn(Selector selector) {
channel = getSock().getChannel();
channel.configureBlocking( false );
channel.register( selector, SelectionKey.OP_WRITE | SelectionKey.OP_READ, this );
outgoing = ByteBuffer.wrap( request.append( "\r\n" ).toString().getBytes() );
}
public boolean isFinished() {
// judge if get "END\r\n"
}
public ByteBuffer getBuffer() {
int last = incoming.size()-1;
if ( last >= 0 && incoming.get( last ).hasRemaining() ) {
return incoming.get( last );
}
else {
ByteBuffer newBuf = ByteBuffer.allocate( 8192 );
incoming.add( newBuf );
return newBuf;
}
}
}
public Object getMulti() throws Exception {
selector = Selector.open();
Conn conn = new Conn(selector);
try {
while(timeRemaining) {
int n = selector.select(timeout));
if ( n > 0 ) {
Iterator<SelectionKey> it = selector.selectedKeys().iterator();
while ( it.hasNext() ) {
SelectionKey key = it.next();
it.remove();
if ( key.isReadable() )
readResponse( key );
else if ( key.isWritable() )
writeRequest( key );
}
}
else {
// error...
}
timeRemaining = timeout - (SystemTimer.currentTimeMillis() - startTime);
}
}
finally {
selector.close();
}
}
public void writeRequest( SelectionKey key ) throws IOException {
ByteBuffer buf = ((Conn) key.attachment()).outgoing;
SocketChannel sc = (SocketChannel)key.channel();
if ( buf.hasRemaining() ) {
sc.write( buf );
}
if ( !buf.hasRemaining() ) {
// switching to read mode for server
key.interestOps( SelectionKey.OP_READ );
}
}
public void readResponse( SelectionKey key ) throws IOException {
Conn conn = (Conn)key.attachment();
ByteBuffer buf = conn.getBuffer();
int count = conn.channel.read( buf );
if ( count > 0 ) {
if ( log.isDebugEnabled() )
log.debug( "read " + count + " from " + conn.channel.socket().getInetAddress() );
if ( conn.isFinished() ) {
...
return;
}
}
}
伪代码中主要给出了NIO中的一些逻辑。并发mget的好处是非常明显的,但这段代码有几个明显的坑。
mget伪代码的几个坑
1. Too many open files的坑
每次getMulti都执行Selector.open()?? Linux系统中,执行Selector.open()打开一对pipe(参考:http://blog.youkuaiyun.com/haoel/article/details/2224055),当后续IO慢时,Selector就不能及时关闭。造成大量pipe被创建,导致Too many open files错误。一般NIO的逻辑是只有一个全局selector,新channel注册后只需selector.wakeup() 即可。
2. 死循环的坑
Java6 NIO有两个众所周知的坑:http://bugs.sun.com/view_bug.do?bug_id=6693490和http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6403933。简单的说,就是Selector应该只在2种情况有返回值,即有网络事件发生或者超时。但是Selector有时却会在没有获得任何selectionKey的情况返回,这是一个Java6 NIO的bug。上面这段mget的伪代码中没有相关处理,容易造成死循环。我们可以参考MINA的解决方法,伪代码如下:
long t0 = System.currentTimeMillis();
int selected = select(1000L);
long t1 = System.currentTimeMillis();
long delta = (t1 - t0);
if ((selected == 0) && !wakeupCalled.get() && (delta < 100)) {
// Last chance : the select() may have been
// interrupted because we have had an closed channel.
if (isBrokenConnection()) {
LOG.warn("Broken connection");
// we can reselect immediately
// set back the flag to false
wakeupCalled.getAndSet(false);
continue;
} else {
LOG.warn("Create a new selector. Selected is 0, delta = " + (t1 - t0));
// Ok, we are hit by the nasty epoll
// spinning.
// Basically, there is a race condition
// which causes a closing file descriptor not to be
// considered as available as a selected channel, but
// it stopped the select. The next time we will
// call select(), it will exit immediately for the same
// reason, and do so forever, consuming 100%
// CPU.
// We have to destroy the selector, and
// register all the socket on a new one.
registerNewSelector();
}
// Set back the flag to false
wakeupCalled.getAndSet(false);
// and continue the loop
continue;
}
这段代码非常清晰,触发条件是selector返回值为0,网络没有断开,并且时间<100ms就认为是触发了Java NIO的bug。处理的方法就是重建一个selector。另外一个可以参考的例子是Jetty的处理方法:http://wiki.eclipse.org/Jetty/Feature/JVM_NIO_Bug