Back in the good old days of the Net, circa the early 1990s, we didn't have the Web and HTTP and graphical browsers. Instead, we had Usenet news and FTP and command-line interfaces, and we liked it that way! But as good as the good old days were, there were some problems. For instance, when we were downloading kilobytes of free software from a popular FTP site over our 2,400 bps modems using Kermit, we would often encounter error messages like this one:

% ftp
Connected to
220 softwarenl FTP server (wu-2.4.2-academ[BETA- 16]+opie-2.32(1) 981105) ready.
Name ( anonymous
530- Server is busy. Please try again later or try one of our other
530- ftp servers at Thank you.
530 User anonymous access denied.
Login failed.

In fact, in the days when the Internet had only a few million users instead of a few hundred million, we were far more likely to come across an overloaded and congested site than we are today. The problem was that both the FTP servers bundled with most Unixes and the third-party FTP servers, such as wu-ftpd, forked a new process for each connection. 100 simultaneous users meant 100 additional processes to handle. Since processes are fairly heavyweight items, too many could rapidly bring a server to its knees. The problem wasn't that the machines weren't powerful enough or the network fast enough; it was that the FTP servers were (and many still are) poorly implemented. Many more simultaneous users could be served if a new process wasn't needed for each connection. Early web servers suffered from this problem as well, although the problem was masked a little by the transitory nature of HTTP connections. Since web pages and their embedded images tend to be small (at least compared to the software archives commonly retrieved by FTP) and since web browsers "hang up" the connection after each file is retrieved instead of staying connected for minutes or hours at a time, web users don't put nearly as much load on a server as FTP users do. However, web server performance still degrades as usage grows. The fundamental problem is that while it's easy to write code that handles each incoming connection and each new task as a separate process (at least on Unix), this solution doesn't scale. By the time a server is attempting to handle a thousand or more simultaneous connections, performance slows to a crawl. There are at least two solutions to this problem. The first is to reuse processes rather than spawning new ones. When the server starts up, a fixed number of processes (say, 300) are spawned to handle requests. Incoming requests are placed in a queue. Each process removes one request from the queue, services the request, then returns to the queue to get the next request. There are still 300 separate processes running, but because all the overhead of building up and tearing down the processes is avoided, these 300 processes can now do the work of 1,000.These numbers are rough estimates. Your exact mileage may vary, especially if your server hasn't yet reached the volume where scalability issues come into play. Still, whatever mileage you get out of spawning new processes, you should be able to do much better by reusing old processes. The second solution to this problem is to use lightweight threads to handle connections instead of heavyweight processes. Whereas each separate process has its own block of memory, threads are easier on resources because they share memory. Using threads instead of processes can buy you another factor of three in server performance. By combining this with a pool of reusable threads (as opposed to a pool of reusable processes), your server can run nine times faster, all on the same hardware and network connection! While it's still the case that most Java virtual machines keel over somewhere between 700 and 2,000 simultaneous threads, the impact of running many different threads on the server hardware is relatively minimal since they all run within one process. Furthermore, by using a thread pool instead of spawning new threads for each connection, a server can use fewer than a hundred threads to handle thousands of connections per minute. Unfortunately, this increased performance doesn't come for free. There's a cost in program complexity. In particular, multithreaded servers (and other multithreaded programs) require programmers to address concerns that aren't issues for single-threaded programs, particularly issues of safety and liveness. Because different threads share the same memory, it's entirely possible for one thread to stomp all over the variables and data structures used by another thread. This is similar to the way one program running on a non-memory-protected operating system such as Mac OS 9 or Windows 95 can crash the entire system. Consequently, different threads have to be extremely careful about which resources they use when. Generally, each thread must agree to use certain resources only when it's sure those resources can't change or that it has exclusive access to them. However, it's also possible for two threads to be too careful, each waiting for exclusive access to resources it will never get. This can lead to deadlock, in which two threads are each waiting for resources the other possesses. Neither thread can proceed without the resources that the other thread has reserved, but neither is willing to give up the resources it has already.


There is a third solution to the problem, which in many cases is the most efficient of all, although it's only available in Java 1.4 and later. Selectors enable one thread to query a group of sockets to find out which ones are ready to be read from or written to, and then process the ready sockets sequentially. In this case, the I/O has to be designed around channels and buffers rather than streams. We'll discuss this in , which demonstrates selector-based solutions to the problems solved in this chapter with threads.