Sometimes the best strategy is to retreat

The Philosopher Developer

February 28, 2011

I have a few remarks to make about my ConcurrentList<T> class, which I have written about a couple of times already.

What, are you “retreating” from your concurrent list idea? In a way, yes. Let me explain. From the beginning, my concept of the usefulness of a lock-free IList<T> implementation was that it would scale better than using a List<T> plus lock statements by eliminating contention when many threads are accessing the list concurrently.

Is this actually true? Accessing the list how?

Since a concurrent implementation of Insert, RemoveAt etc. is not supported in ConcurrentList<T>, there are really only a few operations that need to be considered:

Critically, I think the assumption is that one would need a lock-free data structure to optimize the scenario where multiple threads may be reading from the list (via accessing items at random indexes) while one or more other threads are adding to it. I actually don’t think you need a lock-free data structure for this at all.

Here’s the thing: if you’ve got a List<T>, and you synchronize all calls to Add with, e.g., a lock statement, you don’t need to also synchronize reads. They will always be thread-safe. Accessing an element at an index of a List<T> that is only growing with calls to Add will work from multiple threads. If you are skeptical, consider what actually happens when Add is called (warning: reliance upon specific implementation details ahead!):

  1. The List<T> gets the next available index in its internal array.
  2. If the index is beyond the end of the array:
    1. A new array is allocated.
    2. The elements of the current array are copied over.
    3. A reference to this new array is assigned to the List<T>’s internal array field.
  3. The element in the array at the index found in step 1 is assigned to the input value.

Assuming an always-accurate return value for Count, and assuming the above process is always synchronized, there is no reason why threads reading from the list need to be synchronized.

To demonstrate this, I added the SynchronizedList<T> class to my GitHub repo. Take a look and see for yourself: this class, which simply uses a List<T> internally and synchronizes calls to Add while exposing everything else without synchronization, passes all unit tests.

Now, I told a little fib moments ago. I’m really not giving up on the concurrent list idea; rather, I am rethinking what its actual usefulness is. I still think there is value to having a lock-free Add implementation, which is guaranteed contention-free growth. On a plain List<T> the cost of Add is very small most of the time; but every now and then, it actually requires an expensive array allocation + full copy. In something like a real-time system, this might be a very bad thing to synchronize; the thread contention could create a pretty large albeit ephemeral bottleneck. So maybe the ConcurrentList<T> class can be a very specialized collection for scenarios like this.

I don’t know. To be honest, though I’ve put a lot of thought into this topic, I haven’t put a lot of thought into this blog post specifically; so it’s possible I have not really provided a very compelling or convincing explanation of my feelings on the matter.

What I do know is this: next up, I’m going to propose a data structure which leverages the concept of fast thread-safe adds (whether synchronized via locking or lock-free) to provide O(1) random inserts and removals. Pretty exciting, huh?