The encoding process of LAME on file level is shown in the following picture:

The uncompressed file is read in blocks and the compressed file is written in blocks. Here, read and write operations alternate.

Although this approach and the use of synchronous I/O operations is not optimal the encoding speed of one single file is not affected.

The process can be described in pseudo code in an imperative programming language as follows:

MP3EncodeFile(wavFilename, mp3Filename)
{
  infile = OpenFile(wavFilename);
  outfile = OpenFile(mp3Filename);
  for (i = 0; i < maxBlocks; ++i) {
    inblock = ReadBlock(infile);
    outblock = EncodeBlock(inblock);
    WriteBlock(outfile, outblock);
  }
}

In general, it's possible to call the function "MP3EncodeFile" in parallel on different threads for several files. This is suitable for two or three threads.
With even more threads, too many I/O operations will compete for a serial data device, so that the effiency losses exceed the benefits of more processors.

Fiber Pool has an I/O scheduler to handle such problems. An arbitrary number of tasks can be passed to it that are processed with the best possible I/O performance, then.

The pseudo code in an asynchronous programming language could be as follows:

MP3EncodeFile(wavFilename, mp3Filename)
{
  Memory inMemory;
  Memory outMemory;
  ReadFile(wavFilename, inMemory);
  WriteFile(mp3Filename, outMemory);
  Encode(inMemory, outMemory);
}

By calling this function the three tasks "ReadFile", "WriteFile" and "En­code" are passed to the underlying system (here: Fiber Pool) for further processing.

When processing several files the function can be called arbitrarily, so that for n files 3n tasks will be passed to the system.

Unlike the imperative solution in which the sight distance of a thread is the currently executed program line and no knowledge about further threads exists, the asynchronous system obtains sufficient information in order to develop strategies on how to use CPU, I/O and memory up to the next synchronization point.

Implementation in fpMP3:
The asynchronous function "MP3EncodeFile" was implemented as C++ class "MP3EncodingTask". The I/O strategy was defined to prefer read operations rather than write operations (see documentation of "SetDefaultFileIOSchedulingStrategy").
The CPU and memory strategy can be set through the command line.