TimeSpan and double rounding errors

I'm dealing with physical entities, such as time, speed and distance, and performing simple math, such as distance = time * speed, etc. Speed and Distance are double values rounded to 8th digit, and for Time values a .NET TimeSpan is used. Since TimeSpan rounds to a closest millisecond, I'm getting rounding errors, so I need to write custom rounding method that rounds all calculations to the closest millisecond. For example (rounding to 8th digit is omitted for simplicity):

  static void Main(string[] args) {
     var dist = 1.123451;
     var speed = 1.123452;

     var timeA = TimeSpan.FromHours(dist / speed);
     var timeB = timeA + TimeSpan.FromMilliseconds(1);

     var distA = _round(timeA.TotalHours * speed);
     var distB = _round(timeB.TotalHours * speed);

     var timeA1 = TimeSpan.FromHours(distA / speed);
     var timeB1 = TimeSpan.FromHours(distB / speed);

     // a correct implementation should give both the following vars true
     var isDistributive = distA == dist;
     var isPrecise = (timeB1 - timeA1) == TimeSpan.FromMilliseconds(1);
  }

  public static double _round(double d) {
     // Q: what should be here?
  }

• Using Math.Round(d, 6) is distributive, but loses precision (precise up to ~4 msec)
• Using Math.Round(d, 7) is precise to a single msec, but not distributive (distA above will be 1.1234511 != 1.123451)

• Using the following (round to closest msec) seems to be correct, but the rounding code itself introduces its own double precision errors:

    public static double _round(double d) {
      v开发者_运维知识库ar pre = 3600000.0;
      return Math.Round(d * pre) / pre;
    }
  

Thanks, Boris.

Jon Skeet is likewise unhappy with .Net time calculations, and working on a project called noda-time, which I believe might solve this problem. I don't know if the project is far enough along to be of use to you yet, but it is worth checking.

Edit: Shamelessly invoking the name of JS in the hopes of getting people to use and improve libraries rather than (unnecessarily) reinventing the wheel.

Try using decimal instead of double types. They are more precise (28 digits precision) and should fit your need without having to implement the custom round function.