A faster replacement to the Dictionary<TKey, TValue>

Chances are you're seeing JIT compilation. On my box, I see:

00:00:00.0000360
00:00:00.0000060

when I run it twice in quick succession within the same process - and not in the debugger. (Make sure you're not running it in the debugger, or it's a pointless test.)

Now, measuring any time that tiny is generally a bad idea. You'd need to iterate millions of times to get a better idea of how long it's taking.

Do you have good reason to believe it's actually slowing down your code - or are you basing it all on your original timing?

I doubt that you'll find anything significantly faster than Dictionary<TKey, TValue> and I'd be very surprised to find that it's the bottleneck.

EDIT: I've just benchmarked adding a million elements to a Dictionary<TKey, TValue> where all the keys were existing objects (strings in an array), reusing the same value (as it's irrelevant) and specifying a capacity of a million on construction - and it took about 0.15s on my two-year-old laptop.

Is that really likely to be a bottleneck for you, given that you've already said you're using some "old slow libraries" elsewhere in your app? Bear in mind that the slower those other libraries are, the less impact an improved collection class will have. If the dictionary changes are only accounting for 1% of your overall application time, then even if we could provide an instantaneous dictionary, you'd only speed up your app by 1%.

As ever, get a profiler - it'll give you a much better idea of where your time is going.


I agree with Jon Skeet's supposition that this is most likely JIT compilation.

That being said, I wanted to add some other information here:

Most of the speed issues relating to using Dictionary<T,U> are not related to the implementation of Dictionary. Dictionary<T,U> is VERY fast, out of the box. It would be difficult to beat it.

Speed issues relating to Dictionary instances are almost always actually hash code implementation issues. If you're having speed issues when using Dictionary<MyCustomClass,MyValue>, revisit the GetHashCode() implementation you have defined on MyCustomClass. This is even more critical if you're using a custom struct as your key.

In order to get good performance out of Dictionary, GetHashCode() should be:

  1. Fast
  2. Able to provide hash codes that generate few conflicts. Unique instances should, when possible, generate unique hash values.

If you get that right, I think you'll be very happy with the default Dictionary implementation.


Don't forget, you're timing the Dictionary constructor in that code as well. I did a test, moving the call to the constructor out of the measurement, and looped 10 times. Here's my test code:

for (int i = 0; i < 10; i++)
{
    Dictionary<string, string> test = new Dictionary<string, string>();

    System.Diagnostics.Stopwatch watch = System.Diagnostics.Stopwatch.StartNew();

    test.Add("fieldName", "fieldValue");
    test.Add("Title", "fieldavlkajlkdjflkjalkjslkdjfiajwelkrjelrkjavoijl");

    Console.WriteLine(watch.Elapsed);
}

Console.ReadKey();

Below are the results:

00:00:00.0000607
00:00:00.0000025
00:00:00.0000015
00:00:00.0000015
00:00:00.0000016
00:00:00.0000017
00:00:00.0000016
00:00:00.0000016
00:00:00.0000016
00:00:00.0000015

I'm not sure how much faster you could get than that...

Update

Looks like this mirrors Jon Skeets results too...JIT.