Optimal number of items to keep queued for the thread pool in .NET?

I am building a background processing engine which supports discarding both to-be-processed and is-being-processed items. This is for usage in a winforms application that will require heavy processing of some input elements, so I'm building a queue engine where I can enqueue workload items, and when they're processed, I get notified with the results.

The question is, this queue will almost always contain a lot of items to begin with, and I thought that instead of just dumping everything to the threadpool, I'd place only the first N items into the threadpool, and keep backfilling when they are processed. The reason I want to do this is that once I dump them into the threadpool, they will be processed, and even if they're tagged as discard, they will still take up queue time.

With the backfill implementation I've made, I can remove items from the queue if they become discarded, and only put them into the queue when it's their turn, so to speak.

So the question is, how would I go about calculating this number N, the number of items to place into and keep in the thread pool queue.

Issues I've considered:

• I might want to enqueue 2 * number of processors, which I see is a typical number of items, to ensure all processors are working
• However, if the actual processing of some items is super-fast (which can happen), then the queue in the threadpool is exhausted before my own class can backfill with more work, so perhaps I'd like a bigger number to avoid underutilizing the processors
• Should I create some auto-adjust routine to calculate the optimal number based on the current time each item takes, so that if they are all super-fast, the number is much higher, and if processing takes a bit of time, it should stay low?

What do you think?

New: Ok, due to one of the answers, I'll explain a bit more. Every item put into the queue is keyed by something unique. If I dump another item into the queue with the same key as an existing item, that old item is considered "Discarded", and should be removed. If the item is being processed, a property on the workload item is set to true, a "IsDicarded" property, which the processing method is responsible for calling. If it detects a discarded item, it should quit early, returning no results.

Perhaps I should experiment a bit more, and try to just dump everything into the threadpool.

New question: Is there a limit to the number of items I can queue up? If not, then this would easily simplify my class a lot.

Note: When I say "lengthy processing", I mean in the order of 1-10 seconds. Is the threadpool even the best for this? I see notes all over the web about "the processing should be quick", but what "quick" is is never mentioned. Is quick in the order of milliseconds here?

Do you know Ami Bar's Smart Thread Pool?

Seems its implementation allows you to cancel an unprocessed item and dynamically increases threads as required, until a hard limit; I personally use 100 * Environment.ProcessorsCount