August 15, 2010

Non-blocking operations and deferred execution with node.js

If you write high volume server applications with high concurrency or low latency requirements you have probably heard about node.js This is a relatively easy to understand system that came out in 2009 and has some pretty amazing characteristics. An early presentation by the main author is here - http://s3.amazonaws.com/four.livejournal/20091117/jsconf.pdf


Node.js is an environment for writing Javascript based server applications with a big twist - all IO operations are non-blocking. This non-blocking aspect introduces a concurrency model that may be new to most developers but enables node.js applications to scale to a huge number of concurrent operations - it scales like crazy.

Using non-blocking operations means code that would normally wait for data from a disk file or from a network connection does not wait and waste CPU cycles - your code returns control to the runtime environment and will be called later when the data actually is available. This allows the runtime environment to execute some other code whose data is ready at the moment and gains efficiency by avoiding context switches. This also means there is a single thread accessing data and no synchronization or semaphores are needed to prevent corruption of data due to concurrent access, making your application even more efficient.

Although writing applications in Javascript makes node.js very approachable, the use of non-blocking operations isn't very common in most server applications and results in code that looks similar but is oddly different from what is familiar to most developers. For example, consider a simple program that reads data from a file and processes that data. In a typical procedural program the steps would be :

file = open("filname");
 read(file,buffer);
 close(file);
 do_something(buffer);

This pseudo-code example is easy to understand and probably familiar to most developers. The step-by-step sequence of operations is the way most languages work and how most application logic is described. However, in a non-blocking version the open() function returns immediately - even though the file is not yet open. This introduces some challenges.

file = open("filename");

 // the 'file' is not yet open! what to do?
 read(file,buffer);
 close(file);
 do_something(buffer);

If the open() function were a blocking operation, the runtime environment would defer execution of the remaining sequence of operations until the data was available and then pick up where it left off. In node.js the way that code after a non-blocking operation is paused and picked up later is through the use of callback functions. All the steps listed after using the open() function are bundled into a new function and that bundle of steps is passed as a parameter to the open() function itself. The open() function will return immediately and your code has the choice of doing some work unrelated to the data that is not yet available or simply returning control to the runtime environment by exiting the current function.
When the data for the opened file actually does become available your callback function is invoked by the runtime and your bundle of steps will then proceed.

open("filename",function (f) {
 read(f,buffer);
 close(f);
 do_something(buffer);
});

The parameters to the callback function are defined by the non-blocking operation. In node.js opening files uses a callback that provides an error object (in case opening the file fails) and a file descriptor that can be used to actually read data. In node.js most callback functions have an error object and a list of parameters with the desired data.

In the non-blocking example above you may have noticed the read(f,buffer) function call and guessed that this might be a non-blocking operation. This requires an additional callback function holding the remaining sequence of operations to execute once the data is read into a buffer.

open("filename",function (f) {
 read(f,buffer, function(err,count) {
  close(f);
  do_something(buffer);
 });
});

Some people feel this is a natural way to structure your code. Those people would be wrong.

Here is an actual node.js example of reading from a file

var fs=require('fs'),
 sys=require('sys');


 fs.open("sample.txt",'r',0666,function(err,fd) {
  fs.read(fd,10000,null,'utf8',function(err,str,count) {
   fs.close(fd);
   sys.puts(str);
  });
 });

Although this may appear a bit complex for such a simple task, and you can imagine what happens with more complex application logic, the benefit of this approach becomes more apparent when thinking about more interesting situations. For example, consider reading from two files and merging the contents. Normally a program would read one file, then read another file, then merge the results. The total time taken would be the sum of the time to read each file. With non-blocking operations, reading both files can be started at the same time and the total time taken would only be the longest time to read either of the two files.