"Nested foreach" vs "lambda/linq query" performance(LINQ-to-Objects)

Question

In performance point of view what should you use "Nested foreach's" or "lambda/linq queries"?

Answer 1

Write the clearest code you can, and then benchmark and profile to discover any performance problems. If you do have performance problems, you can experiment with different code to work out whether it's faster or not (measuring all the time with as realistic data as possible) and then make a judgement call as to whether the improvement in performance is worth the readability hit.

A direct foreach approach will be faster than LINQ in many cases. For example, consider:

var query = from element in list
            where element.X > 2
            where element.Y < 2
            select element.X + element.Y;

foreach (var value in query)
{
    Console.WriteLine(value);
}

Now there are two where clauses and a select clause, so every eventual item has to pass through three iterators. (Obviously the two where clauses could be combined in this case, but I'm making a general point.)

Now compare it with the direct code:

foreach (var element in list)
{
    if (element.X > 2 && element.Y < 2)
    {
        Console.WriteLine(element.X + element.Y);
    }
}

That will run faster, because it has fewer hoops to run through. Chances are that the console output will dwarf the iterator cost though, and I'd certainly prefer the LINQ query.

EDIT: To answer about "nested foreach" loops... typically those are represented with SelectMany or a second from clause:

var query = from item in firstSequence
            from nestedItem in item.NestedItems
            select item.BaseCount + nestedItem.NestedCount;

Here we're only adding a single extra iterator, because we'd already be using an extra iterator per item in the first sequence due to the nested foreach loop. There's still a bit of overhead, including the overhead of doing the projection in a delegate instead of "inline" (something I didn't mention before) but it still won't be very different to the nested-foreach performance.

This is not to say you can't shoot yourself in the foot with LINQ, of course. You can write stupendously inefficient queries if you don't engage your brain first - but that's far from unique to LINQ...

Answer 2

If you do

foreach(Customer c in Customer)
{
  foreach(Order o in Orders)
  {
    //do something with c and o
  }
}

You will perform Customer.Count * Order.Count iterations

---

If you do

var query =
  from c in Customer
  join o in Orders on c.CustomerID equals o.CustomerID
  select new {c, o}

foreach(var x in query)
{
  //do something with x.c and x.o
}

You will perform Customer.Count + Order.Count iterations, because Enumerable.Join is implemented as a HashJoin.

Answer 3

It is more complex on that. Ultimately, much of LINQ-to-Objects is (behind the scenes) a foreach loop, but with the added overhead of a little abstraction / iterator blocks / etc. However, unless you do very different things in your two versions (foreach vs LINQ), they should both be O(N).

The real question is: is there a better way of writing your specific algorithm that means that foreach would be inefficient? And can LINQ do it for you?

For example, LINQ makes it easy to hash / group / sort data.

Answer 4

It's been said before, but it merits repeating.

Developers never know where the performance bottleneck is until they run performance tests.

The same is true for comparing technique A to technique B. Unless there is a dramatic difference then you just have to test it. It might be obvious if you have an O(n) vs O(n^x) scenario, but since the LINQ stuff is mostly compiler witchcraft, it merits a profiling.

Besides, unless your project is in production and you have profiled the code and found that that loop is slowing down your execution, leave it as whichever is your preference for readability and maintenance. Premature optimization is the devil.

Answer 5

A great benefit is that using Linq-To-Objects queries gives you the ability to easily turn the query over to PLinq and have the system automatically perform he operation on the correct number of threads for the current system.

If you are using this technique on big datasets, that's an easily become a big win for very little trouble.