Performance of calling delegates vs methods

I haven't seen that effect - I've certainly never encountered it being a bottleneck.

Here's a very rough-and-ready benchmark which shows (on my box anyway) delegates actually being faster than interfaces:

using System;
using System.Diagnostics;

interface IFoo
{
    int Foo(int x);
}

class Program : IFoo
{
    const int Iterations = 1000000000;

    public int Foo(int x)
    {
        return x * 3;
    }

    static void Main(string[] args)
    {
        int x = 3;
        IFoo ifoo = new Program();
        Func<int, int> del = ifoo.Foo;
        // Make sure everything's JITted:
        ifoo.Foo(3);
        del(3);

        Stopwatch sw = Stopwatch.StartNew();        
        for (int i = 0; i < Iterations; i++)
        {
            x = ifoo.Foo(x);
        }
        sw.Stop();
        Console.WriteLine("Interface: {0}", sw.ElapsedMilliseconds);

        x = 3;
        sw = Stopwatch.StartNew();        
        for (int i = 0; i < Iterations; i++)
        {
            x = del(x);
        }
        sw.Stop();
        Console.WriteLine("Delegate: {0}", sw.ElapsedMilliseconds);
    }
}

Results (.NET 3.5; .NET 4.0b2 is about the same):

Interface: 5068
Delegate: 4404

Now I don't have particular faith that that means delegates are really faster than interfaces... but it makes me fairly convinced that they're not an order of magnitude slower. Additionally, this is doing almost nothing within the delegate/interface method. Obviously the invocation cost is going to make less and less difference as you do more and more work per call.

One thing to be careful of is that you're not creating a new delegate several times where you'd only use a single interface instance. This could cause an issue as it would provoke garbage collection etc. If you're using an instance method as a delegate within a loop, you will find it more efficient to declare the delegate variable outside the loop, create a single delegate instance and reuse it. For example:

Func<int, int> del = myInstance.MyMethod;
for (int i = 0; i < 100000; i++)
{
    MethodTakingFunc(del);
}

is more efficient than:

for (int i = 0; i < 100000; i++)
{
    MethodTakingFunc(myInstance.MyMethod);
}

Could this have been the problem you were seeing?

I find it completely implausible that a delegate is substantially faster or slower than a virtual method. If anything the delegate should be negligibly faster. At a lower level, delegates are usually implemented something like (using C-style notation, but please forgive any minor syntax errors as this is just an illustration):

struct Delegate {
    void* contextPointer;   // What class instance does this reference?
    void* functionPointer;  // What method does this reference?
}

Calling a delegate works something like:

struct Delegate myDelegate = somethingThatReturnsDelegate();
// Call the delegate in de-sugared C-style notation.
ReturnType returnValue = 
    (*((FunctionType) *myDelegate.functionPointer))(myDelegate.contextPointer);

A class, translated to C, would be something like:

struct SomeClass {
    void** vtable;        // Array of pointers to functions.
    SomeType someMember;  // Member variables.
}

To call a vritual function, you would do the following:

struct SomeClass *myClass = someFunctionThatReturnsMyClassPointer();
// Call the virtual function residing in the second slot of the vtable.
void* funcPtr = (myClass -> vtbl)[1];
ReturnType returnValue = (*((FunctionType) funcPtr))(myClass);

They're basically the same, except that when using virtual functions you go through an extra layer of indirection to get the function pointer. However, this extra indirection layer is often free because modern CPU branch predictors will guess the address of the function pointer and speculatively execute its target in parallel with looking up the address of the function. I've found (albeit in D, not C#) that virtual function calls in a tight loop are not any slower than non-inlined direct calls, provided that for any given run of the loop they're always resolving to the same real function.