Why ToUpper is faster than ToLower?

Question

Stopwatch stopwatch1 = new Stopwatch();
Stopwatch stopwatch2 = new Stopwatch();
string l = "my test";
string u = "MY TEST";

for (int i = 0; i < 25; i++)
{
    l += l;
    u += u;
}

stopwatch1.Start();
l=l.ToUpper();
stopwatch1.Stop();

stopwatch2.Start();
u=u.ToLower();
stopwatch2.Stop();

// Write result.
Console.WriteLine("Time elapsed: \nUPPER :  {0}\n LOWER : {1}",
                  stopwatch1.Elapsed, stopwatch2.Elapsed);

I have run many times:

UPPER : 00:00:01.3386287
LOWER : 00:00:01.4546552

UPPER : 00:00:01.1614189
LOWER : 00:00:01.1970368

UPPER : 00:00:01.2697430
LOWER : 00:00:01.3460950

UPPER : 00:00:01.2256813
LOWER : 00:00:01.3075738

Answer 1

Let's try reproducing the result

  // Please, notice: the same string for both ToUpper/ToLower
  string GiniPig = string.Concat(Enumerable
    .Range(1, 1000000) // a million chunks "my test/MyTest" combined (long string)
    .Select(item => "my test/MY TEST"));

   Stopwatch sw = new Stopwatch();

   // Let's try n (100) times - not just once
   int n = 100;

   var sampling = Enumerable
     .Range(1, n)
     .Select(x => {
        sw.Reset();
        sw.Start();

        GiniPig.ToLower(); // change this into .ToUpper();

        sw.Stop();
        return sw.ElapsedMilliseconds; })
     .ToSampling(x => x); // Side library; by you may save the data and analyze it with R

   Console.Write(
     $"N = {n}; mean = {sampling.Mean:F0}; std err = {sampling.StandardDeviation:F0}");

Having run several times (warming) I've got the results (Core i7 3.6 GHz, .Net 4.6 IA-64):

ToLower: N = 100; mean = 38; std err = 8
ToUpper: N = 100; mean = 37; std err = 9

So you can't reject the null hypothesis that ToLower is as faster as ToUpper and thus your experiment has got errors:

1. You have different strings to process
2. Processing short (175 characters only) string just once (not in a loop) should be instant and thus the errors can be enourmous
3. You have to warm up the routine (in order methods to be compiled, assemblies loaded, caches filled up etc.)

It seems (the time elapsed is more than 1 second for a very easy operation) it's rule #3 (warming up) breakage which ruined the experiment