Making Code Faster: Going Down the I/O Chute

After introducing the problem and doing some very obvious things, and then doing some pretty non-obvious things, we have managed to get to one-eighth of the initial time of the original implementation.

We can do better still. So far, we relied heavily on the File.ReadLines method, which handles quite a lot of the parsing complexity for us. However, that would still allocate a string per line, and our parsing relied on us splitting the strings again, meaning more allocations.

We can take advantage of our knowledge of the file to do better. The code size blows up, but it is mostly very simple. We create a dedicated record reader class that will read each line of the file with a minimum of allocations.

public class RecordReader : IDisposable
{
    public long Duration;
    public long Id;
    private readonly StreamReader _streamReader;

    private const int SizeOfDate = 19;// 2015-01-01T16:44:31
    private const int SizeOfSpace = 1;
    private const int SizeOfId = 8; // 00043064
    private const int SizeOfNewLine = 2; // \r\n
    private const int SizeOfRecord = SizeOfDate + SizeOfSpace + SizeOfDate + SizeOfSpace + SizeOfId + SizeOfNewLine;

    private readonly char[] _buffer = new char[SizeOfRecord];

    public RecordReader(string file)
    {
        _streamReader = new StreamReader(file);
    }

    public bool MoveNext()
    {
        int sizeRemaining = _buffer.Length;
        int index = 0;
        while (sizeRemaining > 0)
        {
            var read = _streamReader.ReadBlock(_buffer, index, sizeRemaining);
            if (read == 0)
                return false;
            index += read;
            sizeRemaining -= read;
        }

        Duration = (ParseTime(20) - ParseTime(0)).Ticks;
        Id = ParseInt(40, 8);

        return true;
    }

    private DateTime ParseTime(int pos)
    {
        var year = ParseInt(pos, 4);
        var month = ParseInt(pos + 5, 2);
        var day = ParseInt(pos + 8, 2);
        var hour = ParseInt(pos + 11, 2);
        var min = ParseInt(pos + 14, 2);
        var sec = ParseInt(pos + 17, 2);
        return new DateTime(year, month, day, hour, min, sec);
    }

    private int ParseInt(int pos, int size)
    {
        var val = 0;
        for (int i = pos; i < pos + size; i++)
        {
            val *= 10;
            val += _buffer[i] - '0';
        }
        return val;
    }

    public void Dispose()
    {
        _streamReader.Dispose();
    }
}

There is a nontrivial amount of stuff going on here. We start by noting that the size in character of the data is fixed, so we can compute the size of a record very easily. Each record is exactly 50 bytes long.

The key parts here is that we are allocating a single buffer variable, which will hold the line characters. Then we just wrote our own date and integer parsing routines that are very trivial, specific to our case and most importantly, don’t require us to allocate additional strings.

Using this code is done with:

  var stats = new Dictionary<long, FastRecord>();
  using (var reader = new RecordReader(args[0]))
  {
      while (reader.MoveNext())
      {
          FastRecord value;
          if (stats.TryGetValue(reader.Id, out value) == false)
          {
              stats[reader.Id] = value = new FastRecord
              {
                  Id = reader.Id
              };
          }
          value.DurationInTicks += reader.Duration;
      }
  }

So, we are back to single-threaded mode. Running this code gives us a runtime of 1.7 seconds, 126 MB allocated and a peak working set of 35 MB.

We are now about 2.5 times faster than previous parallel version, and over 17 times faster than the original version.

Making this code parallel is fairly trivial now; divide the file into sections and have a record reader on each section. However, is there really much point at this stage?