Which I/O Strategy Should I Use?

by Warren Young

There are several different conventions for communicating with Winsock, and each method has distinct advantages. The question of the hour is, what are these advantages, and how does someone choose the convention that makes the most sense for their application? The choices are:

• Blocking sockets - By default, a Winsock call blocks, meaning that it will not return until it has completed its task or has failed while trying.

• Pure Non-blocking sockets - Calls on non-blocking sockets return immediately, even if they cannot complete their task immediately. Although this allows the program to do other things while the network operations finish, it requires that the program repeatedly poll to find out when each request has finished.

• Asynchronous sockets - These are non-blocking sockets, except that you don't have to poll: the stack sends the program a special window message whenever something "interesting" happens.

• select() - The select() function call is a way to block a thread until something interesting happens on any of a group of sockets. It is usually used with non-blocking sockets, in order to avoid polling.

• Event objects - Used with WSAEventSelect(), this mechanism is similar to the select() method, but a bit more efficient. It also only works on platforms with Winsock, whereas select() works on any platform with BSD sockets.

• Overlapped I/O - One of Winsock 2's major features is that it ties sockets into Win32's unified I/O mechanism. In particular, you can now use overlapped I/O on sockets, which is intrinsically more efficient than the above options.

Further confusing the issue are threads, because each of the above mechanisms changes in nature when used with threads.

In trying to find an answer to the "which I/O strategy" question, it becomes apparent that there are only a few major kinds of programs, and the successful ones follow the same patterns. From those patterns and practical experience — some personal and some borrowed — I have derived the following set of heuristics. None of these heuristics are absolute laws, no one isolated heuristic is sufficient, and the heuristics sometimes conflict. When two heuristics conflict, you need to decide which is more important to your application and ignore the other. However, beware of ignoring a heuristic simply because violating it does not create noticeable consequences for your program. If you get into the habit of ignoring a certain heuristic, it becomes useless.

The heuristics are ordered in terms of compatibility, then speed, and finally functionality. Compatibility is first, because if a given I/O strategy won't work on the platforms you need to support, it doesn't matter how fast or functional it is. Speed is next because performance requirements are easy to determine, and often important. Functionality is last, because once you decide the compatibility and speed issues, your choices become much more subjective.

Heuristic 1: Narrow your choices by deciding which operating systems you need to support.

There are many versions of Windows, but when it comes to the network stack, you can put most of them into one of two groups: the Windows 95 derivatives and the Windows NT 4.0 derivatives. This article treats everything else — Windows NT 3.x, Win16, Windows CE and non-Windows platforms — separately.

Your code may also need to be compatible with POSIX-based systems. This includes Unix, Linux, MacOS X, QNX, and BeOS. Although there are a few different network and threading APIs used by the various POSIX-based systems, I'll only talk about BSD sockets and POSIX threads in this article.

None of these operating systems have exactly the same set of networking features. You can exploit this fact to rule out I/O strategies that not all of your target operating systems support.

1. Win9x does not support overlapped I/O in the kernel. Where overlapped I/O calls work on Win9x, it is because the mechanism is emulated at the API layer. (This applies to Winsock, file and serial/parallel port I/O at least.) This means that programs that only use overlapped I/O functionality guaranteed by the Winsock spec will run fine on Win9x. If, on the other hand, you stray into functionality that only WinNT 4+ provides, your application will fail on Win9x. One example of this is calling ReadFile() with a socket: this works fine on NT4+, but will fail on Win9x.

2. If you only need scatter/gather I/O support, BSD sockets provides this functionality in the readv() and writev() calls. There is no standard Unix mechanism that provides similar efficiencies to Win32's overlapped I/O. Some Unixes provide the aio_*() family of functions (called asynchronous I/O, but not related to Winsock's asynchronous I/O), but this is not implemented widely at the moment.

3. Although all current Unixes support POSIX threads, there are still a lot of older Unix machines out there with broken, nonstandard or nonexistent threading. You will have to choose a subset of all the Unixes if you want to use the same threading code on all Unixes. You'll definitely be writing different threading code for Windows, since its threading API is completely different.

Heuristic 2: Avoid select().

select() is the least efficient way to manage non-blocking I/O, because there is a lot of overhead associated with the function. Most of this overhead is a linear function of the number of connections: double the number of connections, and you double the processing time.

About the only time you should use select() is for compatibility reasons: it's the only non-blocking I/O strategy that works on all versions of Windows (including CE) and on virtually all POSIX-based systems. If your program only needs to work on non-CE versions of Windows, there are better alternatives.

Heuristic 3: Asynchronous sockets work best with low volumes of data.

Asynchronous Winsock I/O (WSAAsyncSelect()) isn't the most efficient I/O strategy, but it's not the least efficient, either. It's a fine way to go in a program that deals with low volumes of data. As the volume of data goes up, the overhead becomes more significant.

Heuristic 4: For high-performance servers, prefer overlapped I/O.

Of all the various I/O strategies, overlapped I/O has the highest performance. (I/O completion ports are even more efficient, but are nonstandard vis-a-vis Winsock proper, so I don't cover them in the FAQ.) With careful use of overlapped I/O (and boatloads of memory in the server!) you can support tens of thousands of connections with a single server. No other I/O strategy comes close to the scalability of overlapped I/O.

Heuristic 5: To support a moderate number of connections, consider asynchronous sockets and event objects.

If your server only has to support a moderate number of connections — up to 100 or so — you may not need overlapped I/O. Overlapped I/O is not easy to program, so if you don't need its efficiencies, you can save yourself a lot of trouble by using a simpler I/O strategy.

Programmed correctly, asynchronous sockets are a reasonable choice for a dedicated server supporting a moderate number of connections. The main problem with doing this is that many servers don't have a user interface, and thus no message loop. A server without a UI using asynchronous sockets would have to create an invisible window solely to support its asynchronous sockets. If your program already has a user interface, though, asynchronous sockets can be the least painful way to add a network server feature to it.

Another reasonable choice for handling a moderate number of connections is event objects. These are very efficient in and of themselves. The main problem you run into with them is that you cannot block on more than 64 event objects at a time. To block on more, you need to create multiple threads, each of which blocks on a subset of the event objects. Before choosing this method, consider that handling 1024 sockets requires 16 threads. Any time you have many more active threads than you have processors in the system, you start causing serious performance problems.

One caution: it's very easy to underestimate the number of simultaneous connections you will get on a public Internet server. It may make sense to design for massive scalability even if your estimates don't currently predict thousands of simultaneous clients.

Heuristic 6: Low-traffic servers can use most any I/O strategy.

For low-traffic servers, there isn't much call to be super-efficient. Perhaps your server just doesn't see high traffic, or perhaps it's running a Windows 95 derivative and so it limited to 100 sockets at a time by the OS. Suitable strategies for 1-100 connections are event objects, non-blocking sockets with select(), asynchronous sockets, and threads with blocking sockets.

We've covered the first three methods already, so let's consider threads with blocking sockets. This is often the simplest way by far to write a server. You just have a main loop that accepts connections and spins each new connection off to its own thread, where it's handled with blocking sockets. Blocking sockets have several advantages. They are efficient, because when a thread blocks, the operating system immediately lets other threads run. Also, synchronous code is more straightforward than equivalent non-synchronous code.

There are two main problems with thread-per-connection servers. First, threads often require a lot of synchronization work, which is hard to get right; this may outstrip the simplicity benefits of using blocking sockets. Second, threads don't scale well at all: as the number of threads increases, the operating system overhead associated with context switches between the threads becomes significant. This method is only suitable for a fairly small number of connections, or a greater number of connections that are mostly idle.

Heuristic 7: Do not block inside a user interface thread.

This heuristic sounds more like a straightforward rule of Windows programming, but I bring it up because most programs are single-threaded. In a single-threaded GUI program, any time you call a Winsock function that blocks the UI thread, buttons can't be pressed, menus won't pull down, scroll bars won't move, keypresses are ignored...your UI freezes.

Heuristic 8: For GUI client programs, prefer asynchronous sockets.

There are two reasons for this heuristic:

1. Asynchronous sockets were designed from the start to work well with GUI programs. You already have a window loop going, and you already have window management code in the rest of the program. Adding asynchronous network I/O is about as easy as adding a dialog to your program.
2. All of the alternatives require at least one additional thread to handle the networking in order to satisfy the previous heuristic. With asynchronous sockets, you can handle both the network and the UI with a single thread. Since window messages are handled one at a time in the order they arrive, everything is automatically synchronized.

Heuristic 9: Threads are rarely helpful in client programs.

When a programmer first learns about threads, he is eager to try them out in his own programs. He sees that they have several advantages, but he doesn't yet see the drawbacks. Unfortunately for the soon-to-be-educated newbie, these drawbacks can have very significant consequences.

One real benefit of threads is that a thread doing I/O on a blocking socket has a linear control flow, and is therefore easier to understand. Asynchronous code is more spread out, so it is harder to write and debug.

Another perceived benefit of threads is a kind of encapsulation: a programmer can split a program up into a number of threads, each of which has a single well-defined task. But, this is only valid if each thread is mostly independent from the rest of the program. If not, the threads will have to share data through a common data structure, destroying any potential encapsulation.

In the end, the biggest problem with threads is also related to shared data structures: synchronization. This issue is covered better elsewhere, so I won't spend many words on it here. In short, synchronization is hard to get right: poorly-synchronized threads are subject to serialization delays, context switching overhead, deadlocks, race conditions and corrupted data. These are hard problems, and for most programs the benefits are not large enough to make them worth overcoming.

A saner alternative is to use asynchronous I/O. This buys you the synchronization benefits described in the previous heuristic. You can even partition the application in a similar manner to threads by creating an invisible window for each socket. If you have two different types of sockets, each socket can have its notifications sent to a different type of window. In straight API terms it means a separate WndProc() for each type of socket. In terms of frameworks like MFC, you can put the code for each type of socket in a different subclass of CWnd.

Heuristic 10: Use threads only when their effect on the rest of the program is easily contained.

The previous heuristic cautions that threads are often very hard to program correctly, but the truth is that they are sometimes very useful. You can make an educated guess about whether threads will improve the program by doing a bit of design work: is there a clean interface between each thread and the rest of the program? If so, synchronization becomes simple. If not, you're going to end up with a mess that crashes and destroys data unpredictably.

Examples where threads are viable are:

1. An FTP server. One way to write an FTP server is to let the main thread accept the incoming network connections, and send each one to a separate thread. Then, each thread can process the incoming FTP commands, send any required replies, and terminate when the session closes. Because each thread never has to interact with any other, and they all act alike, this is an ideal application of threads. (But, keep in mind the previous server-related heuristics: one thread per client severely limits your server's scalability.)
2. A web browser. When you download a file with a modern web browser, the file comes down in the background, so that you can continue browsing. That download stream is most likely handled by a dedicated thread.
3. An email program. In an email program, the primary focus is usually on reading and writing email. However, when an email message needs to be sent, it is best not to interrupt the user's work. You can send that message with a separate network thread, since the process affects the rest of the program only minimally.
4. A stock ticker. Reduced to basics, a stock ticker simply displays a small amount of continuous real-time data in a pleasing and useful format. When the amount of network data involved is low, the thread synchronization overhead becomes negligible. Plus, this kind of application only has a single data structure that needs protection; the really big synchronization problems appear when multiple data structures need to be protected.

Heuristic 11: Design around your protocol.

Some network protocols are inherently synchronous, and others are not. An example of a synchronous protocol is the POP3 e-mail protocol: send a user name, get a response, send a password, get a response, send a request to get the list of emails, get a response... With POP, you have to send these commands in a specific order: you can't send the password before the user name, and you can't get the list of emails without sending the user name and password. Writing a POP client with a non-synchronous socket type would require also writing a state machine.

On the other hand, if your protocol is non-synchronous, you might as well use non-synchronous sockets. Non-synchronous protocols tend to resemble a set of function calls. Consider, for example, a program to retrieve data from a networked SQL database: send a SQL statement, and retrieve the result set. At the end of each "function call", the program is back to its original state: you don't need to maintain a state machine to keep track of where you are in the protocol.

Heuristic 12: Blocking sockets are simpler, non-blocking sockets are more powerful.

This heuristic is almost a restatement of all the above material. It just bears repeating that, while blocking sockets are attractive for their simplicity, you may find that their disadvantages eventually force you to redesign your program to use some form of non-blocking sockets. This is especially true if your program will be supporting more than one socket. (Virtuall all server programs fall into this category.) The only reasonable way to use multiple blocking sockets at once is to use threads, but with non-blocking sockets, you have many more design options.

Conclusion

It is my hope that you find these heuristics helpful. Although you may not agree with each of them, I think that they will at least make you think about your own choices. Design is a highly subjective enterprise, and this list is based mainly on my own thoughts and preferences.

Special thanks go to Philippe Jounin for his comments on the 1998 version of this paper. The 2000 version reflects my greater experience, as well as commentary from David Schwartz and Alun Jones, both of whom expanded my ideas of the proper way to build a Winsock server.

Copyright © 1998-2004 by Warren Young. All rights reserved.