I've an application that makes use of parallelization for processing data.
The main program is in C#, while one of the routine for analyzing data is on an external C++ dll. This library scans data and calls a callback everytime a certain signal is found within the data. Data should be collected, sorted and then stored into HD.
Here is my first simple implementation of the method invoked by the callback and of the method for sorting and storing data:
// collection where saving found signals
List<MySignal> mySignalList = new List<MySignal>();
// method invoked by the callback
private void Collect(int type, long time)
{
lock(locker) { mySignalList.Add(new MySignal(type, time)); }
}
// store signals to disk
private void Store()
{
// sort the signals
mySignalList.Sort();
// file is a object that manages the writing of data to a FileStream
file.Write(mySignalList.ToArray());
}
Data is made up of a bidimensional array (short[][] data) of size 10000 x n, with n variable. I use parallelization in this way:
Parallel.For(0, 10000, (int i) =>
{
// wrapper for the external c++ dll
ProcessData(data[i]);
}
Now for each of the 10000 arrays I estimate that 0 to 4 callbacks could be fired. I'm facing a bottleneck and given that my CPU resources are not over-utilized, I suppose that the lock (together with thousand of callbacks) is the problem (am I right or there could be something else?). I've tried the ConcurrentBag collection but performances are still worse (in line with other user findings).
I thought that a possible solution for use lock-free code would be to have multiple collections. Then it would be necessary a strategy to make each thread of the parallel process working on a single collection. Collections could be for instance inside a dictionary with thread ID as key, but I do not know any .NET facility for this (I should know the threads ID for initialize the dictionary before launching the parallelization). Could be this idea feasible and, in case yes, does exist some .NET tool for this? Or alternatively, any other idea to speed up the process?
[EDIT]
I've followed the Reed Copsey's suggestion and I used the following solution (according to the profiler of VS2010, before the burden for locking and adding to the list was taking 15% of the resources, while now only 1%):
// master collection where saving found signals
List<MySignal> mySignalList = new List<MySignal>();
// thread-local storage of data (each thread is working on its List<MySignal>)
ThreadLocal<List<MySignal>> threadLocal;
// analyze data
private void AnalizeData()
{
using(threadLocal = new ThreadLocal<List<MySignal>>(() =>
{ return new List<MySignal>(); }))
{
Parallel.For<int>(0, 10000,
() =>
{ return 0;},
(i, loopState, localState) =>
{
// wrapper for the external c++ dll
ProcessData(data[i]);
return 0;
},
(localState) =>
{
lock(this)
{
// add thread-local lists to the master collection
mySignalList.AddRange(local.Value);
local.Value.Clear();
}
});
}
}
// method invoked by the callback
private void Collect(int type, long time)
{
local.Value.Add(new MySignal(type, time));
}
thought that a possible solution for use lock-free code would be to have multiple collections. Then it would be necessary a strategy to make each thread of the parallel process working on a single collection. Collections could be for instance inside a dictionary with thread ID as key, but I do not know any .NET facility for this (I should know the threads ID for initialize the dictionary before launching the parallelization). Could be this idea feasible and, in case yes, does exist some .NET tool for this? Or alternatively, any other idea to speed up the process?
You might want to look at using ThreadLocal<T> to hold your collections. This automatically allocates a separate collection per thread.
That being said, there are overloads of Parallel.For which work with local state, and have a collection pass at the end. This, potentially, would allow you to spawn your ProcessData wrapper, where each loop body was working on its own collection, and then recombine at the end. This would, potentially, eliminate the need for locking (since each thread is working on it's own data set) until the recombination phase, which happens once per thread (instead of once per task,ie: 10000 times). This could reduce the number of locks you're taking from ~25000 (0-4*10000) down to a few (system and algorithm dependent, but on a quad core system, probably around 10 in my experience).
For details, see my blog post on aggregating data with Parallel.For/ForEach. It demonstrates the overloads and explains how they work in more detail.
You don't say how much of a "bottleneck" you're encountering. But let's look at the locks.
On my machine (quad core, 2.4 GHz), a lock costs about 70 nanoseconds if it's not contended. I don't know how long it takes to add an item to a list, but I can't imagine that it takes more than a few microseconds. But let's it takes 100 microseconds (I would be very surprised to find that it's even 10 microseconds) to add an item to the list, taking into account lock contention. So if you're adding 40,000 items to the list, that's 4,000,000 microseconds, or 4 seconds. And I would expect one core to be pegged if this were the case.
I haven't used ConcurrentBag, but I've found the performance of BlockingCollection to be very good.
I suspect, though, that your bottleneck is somewhere else. Have you done any profiling?
The basic collections in C# aren't thread safe.
The problem you're having is due to the fact that you're locking the entire collection just to call an add() method.
You could create a thread-safe collection that only locks single elements inside the collection, instead of the whole collection.
Lets look at a linked list for example.
Implement an add(item (or list)) method that does the following:
Lock collection.
A = get last item.
set last item reference to the new item (or last item in new list).
lock last item (A).
unclock collection.
add new items/list to the end of A.
unlock locked item.
This will lock the whole collection for just 3 simple tasks when adding.
Then when iterating over the list, just do a trylock() on each object. if it's locked, wait for the lock to be free (that way you're sure that the add() finished).
In C# you can do an empty lock() block on the object as a trylock().
So now you can add safely and still iterate over the list at the same time.
Similar solutions can be implemented for the other commands if needed.
Any built-in solution for a collection is going to involve some locking. There may be ways to avoid it, perhaps by segregating the actual data constructs being read/written, but you're going to have to lock SOMEWHERE.
Also, understand that Parallel.For() will use the thread pool. While simple to implement, you lose fine-grained control over creation/destruction of threads, and the thread pool involves some serious overhead when starting up a big parallel task.
From a conceptual standpoint, I would try two things in tandem to speed up this algorithm:
Create threads yourself, using the Thread class. This frees you from the scheduling slowdowns of the thread pool; a thread starts processing (or waiting for CPU time) when you tell it to start, instead of the thread pool feeding requests for threads into its internal workings at its own pace. You should be aware of the number of threads you have going at once; the rule of thumb is that the benefits of multithreading are overcome by the overhead when you have more than twice the number of active threads as "execution units" available to execute threads. However, you should be able to architect a system that takes this into account relatively simply.
Segregate the collection of results, by creating a dictionary of collections of results. Each results collection is keyed to some token carried by the thread doing the processing and passed to the callback. The dictionary can have multiple elements READ at one time without locking, and as each thread is WRITING to a different collection within the Dictionary there shouldn't be a need to lock those lists (and even if you did lock them you wouldn't be blocking other threads). The result is that the only collection that has to be locked such that it would block threads is the main dictionary, when a new collection for a new thread is added to it. That shouldn't have to happen often if you're smart about recycling tokens.
Related
I have a C# .NET program that uses an external API to process events for real-time stock market data. I use the API callback feature to populate a ConcurrentDictionary with the data it receives on a stock-by-stock basis.
I have a set of algorithms that each run in a constant loop until a terminal condition is met. They are called like this (but all from separate calling functions elsewhere in the code):
Task.Run(() => ExecutionLoop1());
Task.Run(() => ExecutionLoop2());
...
Task.Run(() => ExecutionLoopN());
Each one of those functions calls SnapTotals():
public void SnapTotals()
{
foreach (KeyValuePair<string, MarketData> kvpMarketData in
new ConcurrentDictionary<string, MarketData>(Handler.MessageEventHandler.Realtime))
{
...
The Handler.MessageEventHandler.Realtime object is the ConcurrentDictionary that is updated in real-time by the external API.
At a certain specific point in the day, there is an instant burst of data that comes in from the API. That is the precise time I want my ExecutionLoop() functions to do some work.
As I've grown the program and added more of those execution loop functions, and grown the number of elements in the ConcurrentDictionary, the performance of the program as a whole has seriously degraded. Specifically, those ExecutionLoop() functions all seem to freeze up and take much longer to meet their terminal condition than they should.
I added some logging to all of the functions above, and to the function that updates the ConcurrentDictionary. From what I can gather, the ExecutionLoop() functions appear to access the ConcurrentDictionary so often that they block the API from updating it with real-time data. The loops are dependent on that data to meet their terminal condition so they cannot complete.
I'm stuck trying to figure out a way to re-architect this. I would like for the thread that updates the ConcurrentDictionary to have a higher priority but the message events are handled from within the external API. I don't know if ConcurrentDictionary was the right type of data structure to use, or what the alternative could be, because obviously a regular Dictionary would not work here. Or is there a way to "pause" my execution loops for a few milliseconds to allow the market data feed to catch up? Or something else?
Your basic approach is sound except for one fatal flaw: they are all hitting the same dictionary at the same time via iterators, sets, and gets. So you must do one thing: in SnapTotals you must iterate over a copy of the concurrent dictionary.
When you iterate over Handler.MessageEventHandler.Realtime or even new ConcurrentDictionary<string, MarketData>(Handler.MessageEventHandler.Realtime) you are using the ConcurrentDictionary<>'s iterator, which even though is thread-safe, is going to be using the dictionary for the entire period of iteration (including however long it takes to do the processing for each and every entry in the dictionary). That is most likely where the contention occurs.
Making a copy of the dictionary is much faster, so should lower contention.
Change SnapTotals to
public void SnapTotals()
{
var copy = Handler.MessageEventHandler.Realtime.ToArray();
foreach (var kvpMarketData in copy)
{
...
Now, each ExecutionLoopX can execute in peace without write-side contention (your API updates) and without read-side contention from the other loops. The write-side can execute without read-side contention as well.
The only "contention" should be for the short duration needed to do each copy.
And by the way, the dictionary copy (an array) is not threadsafe; it's just a plain array, but that is ok because each task is executing in isolation on its own copy.
I think that your main problem is not related to the ConcurrentDictionary, but to the large number of ExecutionLoopX methods. Each of these methods saturates a CPU core, and since the methods are more than the cores of your machine, the whole CPU is saturated. My assumption is that if you find a way to limit the degree of parallelism of the ExecutionLoopX methods to a number smaller than the Environment.ProcessorCount, your program will behave and perform better. Below is my suggestion for implementing this limitation.
The main obstacle is that currently your ExecutionLoopX methods are monolithic: they can't be separated to pieces so that they can be parallelized. My suggestion is to change their return type from void to async Task, and place an await Task.Yield(); inside the outer loop. This way it will be possible to execute them in steps, with each step being the code from the one await to the next.
Then create a TaskScheduler with limited concurrency, and a TaskFactory that uses this scheduler:
int maxDegreeOfParallelism = Environment.ProcessorCount - 1;
TaskScheduler scheduler = new ConcurrentExclusiveSchedulerPair(
TaskScheduler.Default, maxDegreeOfParallelism).ConcurrentScheduler;
TaskFactory taskFactory = new TaskFactory(scheduler);
Now you can parallelize the execution of the methods, by starting the tasks with the taskFactory.StartNew method instead of the Task.Run:
List<Task> tasks = new();
tasks.Add(taskFactory.StartNew(() => ExecutionLoop1(data)).Unwrap());
tasks.Add(taskFactory.StartNew(() => ExecutionLoop2(data)).Unwrap());
tasks.Add(taskFactory.StartNew(() => ExecutionLoop3(data)).Unwrap());
tasks.Add(taskFactory.StartNew(() => ExecutionLoop4(data)).Unwrap());
//...
Task.WaitAll(tasks.ToArray());
The .Unwrap() is needed because the taskFactory.StartNew returns a nested task (Task<Task>). The Task.Run method is also doing this unwrapping internally, when the action is asynchronous.
An online demo of this idea can be found here.
The Environment.ProcessorCount - 1 configuration means that one CPU core will be available for other work, like the communication with the external API and the updating of the ConcurrentDictionary.
A more cumbersome implementation of the same idea, using iterators and the Parallel.ForEach method instead of async/await, can be found in the first revision of this answer.
If you're not squeamish about mixing operations in a task, you could redesign such that instead of task A doing A things, B doing B things, C doing C things, etc. you can reduce the number of tasks to the number of processors, and thus run fewer concurrently, greatly easing contention.
So, for example, say you have just two processors. Make a "general purpose/pluggable" task wrapper that accepts delegates. So, wrapper 1 would accept delegates to do A and B work. Wrapper 2 would accept delegates to do C and D work. Then ask each wrapper to spin up a task that calls the delegates in a loop over the dictionary.
This would of course need to be measured. What I am proposing is, say, 4 tasks each doing 4 different types of processing. This is 4 units of work per loop over 4 loops. This is not the same as 16 tasks each doing 1 unit of work. In that case you have 16 loops.
16 loops intuitively would cause more contention than 4.
Again, this is a potential solution that should be measured. There is one drawback for sure: you will have to ensure that a piece of work within a task doesn't affect any of the others.
I have a program that is constantly reading and parsing a large stream of data from a WebSocket. All of the parsing happens on one thread within the client, and the data is organized into a SortedSet<T> tree for fast operation.
All of the data is added, updated, and removed without a hitch.
The problem comes when I try to access the data from another thread. It will run fine, but somewhere along the lines is a race condition that will be hit within a minute or two.
Consider this code (running on its own thread) to update the UI in near real-time:
private async Task RenderOrderBook()
{
var book = _client.OrderBook;
while (true)
{
try
{
var asks = book.Asks.OrderBy(i => i.Price).Take(5).OrderByDescending(i => i.Price);
var bids = book.Bids.OrderByDescending(i => i.Price).Take(5);
orderBookView.BeginInvoke(new MethodInvoker(() =>
{
...omitted due to irrelevance
}));
await Task.Delay(500);
}
catch (Exception ex)
{
ex.ToString();
}
}
}
The race condition lies within the LINQ operations on book. The common error is that i.Price (a decimal variable), or perhaps just the object i is referring to, is null. Additionally, my shoddy attempt to just swallow the exception does not actually work.
Regardless, my guess is that the data is being parsed and manipulated so fast that eventually, when using the LINQ OrderBy operation, it will hit a case where a node has been removed by the client, attempt to read from it, and throw an exception.
The book.Asks and book.Bids properties were initially of type SortedSet<T> and pointed directly to the data member itself. In an attempt to mitigate this race condition scenario, I attempted to change them to an array of the node, and use a _asks.ToArray() call to essentially make a copy to read from. This helped make the problem occur a bit less frequently, but nonetheless it still does happen.
How can I make this thread-safe?
Additional Code Snippets
public PriceNode[] Asks
{
get { return _asks.ToArray(); }
}
public PriceNode[] Bids
{
get { return _bids.ToArray(); }
}
My first rule of UI development is that you never perform I/O on the UI thread. Sounds like you've got that one covered.
My second rule is that once something is visible to the UI thread, you can't touch it from any other thread. There is exactly one exception to this rule, and that is for immutable data: if an object will not change, then any thread can touch it. Mutable data? No touch. Keep in mind that "mutable data" includes most collections.
Your life will be so much easier if you can follow these two rules. Following one without breaking the other can be tricky, but there are ways to do it, and once you have a decent grip of them, you'll be in a better place. The path to enlightenment begins here:
Your read thread (the thread reading off the socket) is allowed to create all the new objects it wants, but it can't update existing objects. It also can't modify any collections that the UI thread is using. If you're only adding new objects, this isn't so bad: your read thread can pull data off the socket and use it to cook up new objects. When those objects are ready, it has to hand them over to the UI thread, and the UI thread can add them to the relevant collections. The bulk of the work (and all of the I/O) happens on the read thread, which is what we want, per Strobel's Rule #1. The act of "committing" the already-populated objects should be trivial by comparison. Per Rule #2, once any mutable objects get handed off to the UI thread, your read thread can't touch them again. Ever.
Updating existing objects is trickier. There's a couple ways you can approach this. One is to have the read thread use the latest data to create new objects, which it then hands off to the UI thread. If you have very simple object graphs, the easiest option might be to simply replace the old objects with their newer versions, keeping in mind that any UI code referencing an old object will need to know that it's been replaced. Alternatively, the UI thread can use the data from the new object to update the existing object. If you're following Rule #2, this will be totally thread-safe, and any UI code that pointed to the old object automatically sees the new data without any torn reads or other race-related nastiness. This approach is probably your best bet.
If, after trying out the approaches in the previous paragraph, you find that you are generating unacceptable amounts of garbage, there is a third option. The read thread can copy the raw data for each object into a temporary buffer, then hand the buffers over to the UI thread, which can use the data in the buffers to update the existing objects. This means more work occurring on the UI thread, but at least the data is already in memory (the socket I/O is already done). Since the point of this approach is to create less garbage, it only makes sense if you reuse the buffers. That means you need a thread-safe buffer pool. The read thread acquires a temporary buffer, fills it from the socket, hands it to the UI thread, which returns it to the pool when it's done. Astute readers will note that passing mutable buffers between threads bumps up against Rule #2, so take care that once a thread hands over a buffer, it immediately forgets about it. Because this approach requires a stronger grasp of thread safety to make the pool work, I recommend it only as a last resort. If you can get away with one of the options in the previous paragraph, please do so.
Regardless of which approach you use for updating existing objects, you'll need a way to match up the new objects/data with the old objects. If each object has a unique identifier, you can use a Dictionary<,> as an efficient lookup mechanism. Replacing old objects with their newer copies is a bit more involved, because the old versions may be scattered across multiple collections, some of which may not support efficient replacement.
One last thing: when you hand over new/updated objects to the UI thread, it is vastly preferable to do it in batches. For example, you're better off posting a single operation to your UI thread to update 100 objects than posting 100 separate operations that each update one object.
I've created a game emulation program using c# async socks. I need to remove/add & do iterations on a collection (a list that holds clients) concurrently. I am currently using "lock", however, it's a a huge performance drop. I also do not want to use "local lists/copies" to keep the list up-to-date. I've heard about "ConcurrentBags", however, I am not sure how thread safe they are for iterations (for instance if a thread removes an element from the list while another thread is doing an iteration on it!?).
What do you suggest?
Edit: here is a situation
this is when a packet is sent to all the users in a room
lock (parent.gameClientList)
{
for (int i = 0; i <= parent.gameClientList.Count() - 1; i++) if (parent.gameClientList[i].zoneId == zoneId) parent.gameClientList[i].SendXt(packetElements); //if room matches - SendXt sends a packet
}
When a new client connects
Client connectedClient = new Client(socket, this);
lock (gameClientList)
{
gameClientList.Add(connectedClient);
}
Same case when a client disconnects.
I am asking for a better alternative (performance-wise) because the locks slow down everything.
It sounds like the problem is that you're doing all the work within your foreach loop, and it's locking out the add/remove methods for too long. The way around this is to quickly make a copy of the collection while it's locked, and then you can close the lock and iterate on the copy.
Thing[] copy;
lock(myLock) {
copy = _collection.ToArray();
}
foreach(var thing in copy) {...}
The drawback is that by the time you get around to operating on some object of that copy, it may have been removed from the original collection and so maybe you don't want to operate on it anymore. That's another thing you'll just have to figure out the requirements. If that's a problem, a simple option would be to lock each iteration of the loop, which of course will slow things down but at least it won't lock for the entire duration the loop is running:
foreac(var thing in copy) {
lock(myLock) {
if (_collection.Contains(thing)) //check that it's still in the original colleciton
DoWork(thing); //If you can move this outside the lock it'd make your app snappier, but it could also cause problems if you're doing something "dangerous" in DoWork.
}
}
If this is what you meant by "local copies", then you can disregard this option, but I figured I'd offer it in case you meant something else.
Every time you do something concurrently you are going to have loss due to task management (i.e. locks). I suggest you look at what is the bottleneck in your process. You seem to have a shared memory model, as opposed to a message passing model. If you know you need to modify the entire collection at once, there may not be a good solution. But if you are making changes in a particular order you can leverage that order to prevent delays. Locks is an implementation of pessimistic concurrency. You could switch to an optimistic concurrency model. In one the cost is waiting in the other the cost is retrying. Again the actual solution depends on your use case.
On problem with ConcurrentBag is that it is unordered so you cannot pull items out by index the same way you are doing it currently. However, you can iterate it via foreach to get the same effect. This iteration will be thread-safe. It will not go bizerk if an item is added or removed while the iteration is happening.
There is another problem with ConcurrentBag though. It actually copies the contents to a new List internally to make the enumerator work correctly. So even if you wanted to just pick off a single item via the enumerator it would still be a O(n) operation because of the way enumerator works. You can verify this by disassembling it.
However, based on context clues from your update I assume that this collection is going to be small. It appears that there is only one entry per "game client" which means it is probably going to store a small number of items right? If that is correct then the performance of the GetEnumerator method will be mostly insignificant.
You should also consider ConcurrentDictionary as well. I noticed that you are trying to match items from the collection based on zoneId. If you store the items in the ConcurrentDictionary keyed by zoneId then you would not need to iterate the collection at all. Of course, this assumes that there is only one entry per zoneId which may not be the case.
You mentioned that you did not want to use "local lists/copies", but you never said why. I think you should reconsider this for the following reasons.
Iterations could be lock-free.
Adding and removing items appears to be infrequent based context clues from your code.
There are a couple of patterns you can use to make the list copying strategy work really well. I talk about them in my answers here and here.
I have several actions that I want to execute in the background, but they have to be executed synchronously one after the other.
I was wondering if it's a good idea to use the Task.ContinueWith method to achieve this. Do you foresee any problems with this?
My code looks something like this:
private object syncRoot =new object();
private Task latestTask;
public void EnqueueAction(System.Action action)
{
lock (syncRoot)
{
if (latestTask == null)
latestTask = Task.Factory.StartNew(action);
else
latestTask = latestTask.ContinueWith(tsk => action());
}
}
There is one flaw with this, which I recently discovered myself because I am also using this method of ensuring tasks execute sequentially.
In my application I had thousands of instances of these mini-queues and quickly discovered I was having memory issues. Since these queues were often idle I was holding onto the last completed task object for a long time and preventing garbage collection. Since the result object of the last completed task was often over 85,000 bytes it was allocated to Large Object Heap (which does not perform compaction during garbage collection). This resulted in fragmentation of the LOH and the process continuously growing in size.
As a hack to avoid this, you can schedule a no-op task right after the real one within your lock. For a real solution, I will need to move to a different method of controlling the scheduling.
This should work as designed (using the fact that TPL will schedule the continuation immediately if the corresponding task already has completed).
Personally in this case I would just use a dedicated thread using a concurrent queue (ConcurrentQueue) to draw tasks from - this is more explicit but easier to parse reading the code, especially if you want to find out i.e. how many tasks are currently queued etc.
I used this snippet and have seem to get it work as designed.
The number of instances in my case does not runs in to thousands, but in single digit.
Nevertheless, no issues so far.
I would be interested in the ConcurrentQueue example, if there is any?
Thanks
Greetings.
I'm trying to implement some multithreaded code in an application. The purpose of this code is to validate items that the database gives it. Validation can take quite a while (a few hundred ms to a few seconds), so this process needs to be forked off into its own thread for each item.
The database may give it 20 or 30 items a second in the beginning, but that begins to decline rapidly, eventually reaching about 65K items over 24 hours, at which point the application exits.
I'd like it if anyone more knowledgeable could take a peek at my code and see if there's any obvious problems. No one I work with knows multithreading, so I'm really just on my own, on this one.
Here's the code. It's kinda long but should be pretty clear. Let me know if you have any feedback or advice. Thanks!
public class ItemValidationService
{
/// <summary>
/// The object to lock on in this class, for multithreading purposes.
/// </summary>
private static object locker = new object();
/// <summary>Items that have been validated.</summary>
private HashSet<int> validatedItems;
/// <summary>Items that are currently being validated.</summary>
private HashSet<int> validatingItems;
/// <summary>Remove an item from the index if its links are bad.</summary>
/// <param name="id">The ID of the item.</param>
public void ValidateItem(int id)
{
lock (locker)
{
if
(
!this.validatedItems.Contains(id) &&
!this.validatingItems.Contains(id)
){
ThreadPool.QueueUserWorkItem(sender =>
{
this.Validate(id);
});
}
}
} // method
private void Validate(int itemId)
{
lock (locker)
{
this.validatingItems.Add(itemId);
}
// *********************************************
// Time-consuming routine to validate an item...
// *********************************************
lock (locker)
{
this.validatingItems.Remove(itemId);
this.validatedItems.Add(itemId);
}
} // method
} // class
The thread pool is a convenient choice if you have light weight sporadic processing that isn't time sensitive. However, I recall reading on MSDN that it's not appropriate for large scale processing of this nature.
I used it for something quite similar to this and regret it. I took a worker-thread approach in subsequent apps and am much happier with the level of control I have.
My favorite pattern in the worker-thread model is to create a master thread which holds a queue of tasks items. Then fork a bunch of workers that pop items off that queue to process. I use a blocking queue so that when there are no items the process, the workers just block until something is pushed onto the queue. In this model, the master thread produces work items from some source (db, etc.) and the worker threads consume them.
I second the idea of using a blocking queue and worker threads. Here is a blocking queue implementation that I've used in the past with good results:
https://www.codeproject.com/Articles/8018/Bounded-Blocking-Queue-One-Lock
What's involved in your validation logic? If its mainly CPU bound then I would create no more than 1 worker thread per processor/core on the box. This will tell you the number of processors:
Environment.ProcessorCount
If your validation involves I/O such as File Access or database access then you could use a few more threads than the number of processors.
Be careful, QueueUserWorkItem might fail
There is a possible logic error in the code posted with the question, depending on where the item id in ValidateItem(int id) comes from. Why? Because although you correctly lock your validatingItems and validatedItems queues before queing a work item, you do not add the item to the validatingItems queue until the new thread spins up. That means there could be a time gap where another thread calls ValidateItem(id) with the same id (unless this is running on a single main thread).
I would add item to the validatingItems queue just before queuing the item, inside the lock.
Edit: also QueueUserWorkItem() returns a bool so you should use the return value to make sure the item was queued and THEN add it to the validatingItems queue.
ThreadPool may not be optimal for jamming so much at once into it. You may want to research the upper limits of its capabilities and/or roll your own.
Also, there is a race condition that exists in your code, if you expect no duplicate validations. The call to
this.validatingItems.Add(itemId);
needs to happen in the main thread (ValidateItem), not in the thread pool thread (Validate method). This call should occur a line before the queueing of the work item to the pool.
A worse bug is found by not checking the return of QueueUserWorkItem. Queueing can fail, and why it doesn't throw an exception is a mystery to us all. If it returns false, you need to remove the item that was added to the validatingItems list, and handle the error (throw exeception probably).
I would be concerned about performance here. You indicated that the database may give it 20-30 items per second and an item could take up to a few seconds to be validated. That could be quite a large number of threads -- using your metrics, worst case 60-90 threads! I think you need to reconsider the design here. Michael mentioned a nice pattern. The use of the queue really helps keep things under control and organized. A semaphore could also be employed to control number of threads created -- i.e. you could have a maximum number of threads allowed, but under smaller loads, you wouldn't necessarily have to create the maximum number if fewer ended up getting the job done -- i.e. your own pool size could be dynamic with a cap.
When using the thread-pool, I also find it more difficult to monitor the execution of threads from the pool in their performing the work. So, unless it's fire and forget, I am in favor of more controlled execution. I know you mentioned that your app exits after the 65K items are all completed. How are you monitoring you threads to determine if they have completed their work -- i.e. all queued workers are done. Are you monitoring the status of all items in the HashSets? I think by queuing your items up and having your own worker threads consume off that queue, you can gain more control. Albeit, this can come at the cost of more overhead in terms of signaling between threads to indicate when all items have been queued allowing them to exit.
You could also try using the CCR - Concurrency and Coordination Runtime. It's buried inside Microsoft Robotics Studio, but provides an excellent API for doing this sort of thing.
You'd just need to create a "Port" (essentially a queue), hook up a receiver (method that gets called when something is posted to it), and then post work items to it. The CCR handles the queue and the worker thread to run it on.
Here's a video on Channel9 about the CCR.
It's very high-performance and is even being used for non-Robotics stuff (Myspace.com uses it behind the scenese for their content-delivery network).
I would recommend looking into MSDN: Task Parallel Library - DataFlow. You can find examples of implementing Producer-Consumer in your case would be the database producing items to validate and the validation routine becomes the consumer.
Also recommend using ConcurrentDictionary<TKey, TValue> as a "Concurrent" hash set where you just populate the keys with no values :). You can potentially make your code lock-free.