As we had known, If we want to use traditional IO to construct server, it must block somewhere, so we had to use loop or one thread one socket mode, So nio seem it is better
IMHO, Blocking IO is generally the simplest to use, and unless you have a specific requirement which demands more from your system, you should stick with simplest option.
The next simplest option is blocking NIO, which I often prefer if I want something more efficiency or control than IO. It is still relatively simple but allows you to use ByteBuffers. e.g. ByteBuffers support little endian.
A common option is to use non-blocking NIO with Selectors. Much of the complexity this introduces can be handled by frameworks such as Netty or Mina. I suggest you use such a library if you need non-blocking IO e.g. because you have thousands of concurrent connections per server. IMHO You have thousands of connections, you should consider having more servers unless what each connection does is pretty trivial. AFAIK google go for more servers rather thousands of users per server.
The more extreme option is to use NIO2. This is even more complex and lengthy it write than non-blocking NIO. I don't know of any frameworks which support this well. i.e. it is actually faster when you do. AFAIK It appears this is worth using if you have Infiniband (which is what it was designed to support) but perhaps not worth using if you have Ethernet.