This question is a little trickier than one might expect due to several unknowns: The behaviour of the resource being pooled, the expected/required lifetime of objects, the real reason that the pool is required, etc. Typically pools are special-purpose - thread pools, connection pools, etc. - because it is easier to optimize one when you know exactly what the resource does and more importantly have control over how that resource is implemented.
Since it's not that simple, what I've tried to do is offer up a fairly flexible approach that you can experiment with and see what works best. Apologies in advance for the long post, but there is a lot of ground to cover when it comes to implementing a decent general-purpose resource pool. and I'm really only scratching the surface.
A general-purpose pool would have to have a few main "settings", including:
- Resource loading strategy - eager or lazy;
- Resource loading mechanism - how to actually construct one;
- Access strategy - you mention "round robin" which is not as straightforward as it sounds; this implementation can use a circular buffer which is similar, but not perfect, because the pool has no control over when resources are actually reclaimed. Other options are FIFO and LIFO; FIFO will have more of a random-access pattern, but LIFO makes it significantly easier to implement a Least-Recently-Used freeing strategy (which you said was out of scope, but it's still worth mentioning).
For the resource loading mechanism, .NET already gives us a clean abstraction - delegates.
private Func<Pool<T>, T> factory;
Pass this through the pool's constructor and we're about done with that. Using a generic type with a new() constraint works too, but this is more flexible.
Of the other two parameters, the access strategy is the more complicated beast, so my approach was to use an inheritance (interface) based approach:
public class Pool<T> : IDisposable
{
// Other code - we'll come back to this
interface IItemStore
{
T Fetch();
void Store(T item);
int Count { get; }
}
}
The concept here is simple - we'll let the public Pool class handle the common issues like thread-safety, but use a different "item store" for each access pattern. LIFO is easily represented by a stack, FIFO is a queue, and I've used a not-very-optimized-but-probably-adequate circular buffer implementation using a List<T> and index pointer to approximate a round-robin access pattern.
All of the classes below are inner classes of the Pool<T> - this was a style choice, but since these really aren't meant to be used outside the Pool, it makes the most sense.
class QueueStore : Queue<T>, IItemStore
{
public QueueStore(int capacity) : base(capacity)
{
}
public T Fetch()
{
return Dequeue();
}
public void Store(T item)
{
Enqueue(item);
}
}
class StackStore : Stack<T>, IItemStore
{
public StackStore(int capacity) : base(capacity)
{
}
public T Fetch()
{
return Pop();
}
public void Store(T item)
{
Push(item);
}
}
These are the obvious ones - stack and queue. I don't think they really warrant much explanation. The circular buffer is a little more complicated:
class CircularStore : IItemStore
{
private List<Slot> slots;
private int freeSlotCount;
private int position = -1;
public CircularStore(int capacity)
{
slots = new List<Slot>(capacity);
}
public T Fetch()
{
if (Count == 0)
throw new InvalidOperationException("The buffer is empty.");
int startPosition = position;
do
{
Advance();
Slot slot = slots[position];
if (!slot.IsInUse)
{
slot.IsInUse = true;
--freeSlotCount;
return slot.Item;
}
} while (startPosition != position);
throw new InvalidOperationException("No free slots.");
}
public void Store(T item)
{
Slot slot = slots.Find(s => object.Equals(s.Item, item));
if (slot == null)
{
slot = new Slot(item);
slots.Add(slot);
}
slot.IsInUse = false;
++freeSlotCount;
}
public int Count
{
get { return freeSlotCount; }
}
private void Advance()
{
position = (position + 1) % slots.Count;
}
class Slot
{
public Slot(T item)
{
this.Item = item;
}
public T Item { get; private set; }
public bool IsInUse { get; set; }
}
}
I could have picked a number of different approaches, but the bottom line is that resources should be accessed in the same order that they were created, which means that we have to maintain references to them but mark them as "in use" (or not). In the worst-case scenario, only one slot is ever available, and it takes a full iteration of the buffer for every fetch. This is bad if you have hundreds of resources pooled and are acquiring and releasing them several times per second; not really an issue for a pool of 5-10 items, and in the typical case, where resources are lightly used, it only has to advance one or two slots.
Remember, these classes are private inner classes - that is why they don't need a whole lot of error-checking, the pool itself restricts access to them.
Throw in an enumeration and a factory method and we're done with this part:
// Outside the pool
public enum AccessMode { FIFO, LIFO, Circular };
private IItemStore itemStore;
// Inside the Pool
private IItemStore CreateItemStore(AccessMode mode, int capacity)
{
switch (mode)
{
case AccessMode.FIFO:
return new QueueStore(capacity);
case AccessMode.LIFO:
return new StackStore(capacity);
default:
Debug.Assert(mode == AccessMode.Circular,
"Invalid AccessMode in CreateItemStore");
return new CircularStore(capacity);
}
}
The next problem to solve is loading strategy. I've defined three types:
public enum LoadingMode { Eager, Lazy, LazyExpanding };
The first two should be self-explanatory; the third is sort of a hybrid, it lazy-loads resources but doesn't actually start re-using any resources until the pool is full. This would be a good trade-off if you want the pool to be full (which it sounds like you do) but want to defer the expense of actually creating them until first access (i.e. to improve startup times).
The loading methods really aren't too complicated, now that we have the item-store abstraction:
private int size;
private int count;
private T AcquireEager()
{
lock (itemStore)
{
return itemStore.Fetch();
}
}
private T AcquireLazy()
{
lock (itemStore)
{
if (itemStore.Count > 0)
{
return itemStore.Fetch();
}
}
Interlocked.Increment(ref count);
return factory(this);
}
private T AcquireLazyExpanding()
{
bool shouldExpand = false;
if (count < size)
{
int newCount = Interlocked.Increment(ref count);
if (newCount <= size)
{
shouldExpand = true;
}
else
{
// Another thread took the last spot - use the store instead
Interlocked.Decrement(ref count);
}
}
if (shouldExpand)
{
return factory(this);
}
else
{
lock (itemStore)
{
return itemStore.Fetch();
}
}
}
private void PreloadItems()
{
for (int i = 0; i < size; i++)
{
T item = factory(this);
itemStore.Store(item);
}
count = size;
}
The size and count fields above refer to the maximum size of the pool and the total number of resources owned by the pool (but not necessarily available), respectively. AcquireEager is the simplest, it assumes that an item is already in the store - these items would be preloaded at construction, i.e. in the PreloadItems method shown last.
AcquireLazy checks to see if there are free items in the pool, and if not, it creates a new one. AcquireLazyExpanding will create a new resource as long as the pool hasn't reached its target size yet. I've tried to optimize this to minimize locking, and I hope I haven't made any mistakes (I have tested this under multi-threaded conditions, but obviously not exhaustively).
You might be wondering why none of these methods bother checking to see whether or not the store has reached the maximum size. I'll get to that in a moment.
Now for the pool itself. Here is the full set of private data, some of which has already been shown:
private bool isDisposed;
private Func<Pool<T>, T> factory;
private LoadingMode loadingMode;
private IItemStore itemStore;
private int size;
private int count;
private Semaphore sync;
Answering the question I glossed over in the last paragraph - how to ensure we limit the total number of resources created - it turns out that the .NET already has a perfectly good tool for that, it's called Semaphore and it's designed specifically to allow a fixed number of threads access to a resource (in this case the "resource" is the inner item store). Since we're not implementing a full-on producer/consumer queue, this is perfectly adequate for our needs.
The constructor looks like this:
public Pool(int size, Func<Pool<T>, T> factory,
LoadingMode loadingMode, AccessMode accessMode)
{
if (size <= 0)
throw new ArgumentOutOfRangeException("size", size,
"Argument 'size' must be greater than zero.");
if (factory == null)
throw new ArgumentNullException("factory");
this.size = size;
this.factory = factory;
sync = new Semaphore(size, size);
this.loadingMode = loadingMode;
this.itemStore = CreateItemStore(accessMode, size);
if (loadingMode == LoadingMode.Eager)
{
PreloadItems();
}
}
Should be no surprises here. Only thing to note is the special-casing for eager loading, using the PreloadItems method already shown earlier.
Since almost everything's been cleanly abstracted away by now, the actual Acquire and Release methods are really very straightforward:
public T Acquire()
{
sync.WaitOne();
switch (loadingMode)
{
case LoadingMode.Eager:
return AcquireEager();
case LoadingMode.Lazy:
return AcquireLazy();
default:
Debug.Assert(loadingMode == LoadingMode.LazyExpanding,
"Unknown LoadingMode encountered in Acquire method.");
return AcquireLazyExpanding();
}
}
public void Release(T item)
{
lock (itemStore)
{
itemStore.Store(item);
}
sync.Release();
}
As explained earlier, we're using the Semaphore to control concurrency instead of religiously checking the status of the item store. As long as acquired items are correctly released, there's nothing to worry about.
Last but not least, there's cleanup:
public void Dispose()
{
if (isDisposed)
{
return;
}
isDisposed = true;
if (typeof(IDisposable).IsAssignableFrom(typeof(T)))
{
lock (itemStore)
{
while (itemStore.Count > 0)
{
IDisposable disposable = (IDisposable)itemStore.Fetch();
disposable.Dispose();
}
}
}
sync.Close();
}
public bool IsDisposed
{
get { return isDisposed; }
}
The purpose of that IsDisposed property will become clear in a moment. All the main Dispose method really does is dispose the actual pooled items if they implement IDisposable.
Now you can basically use this as-is, with a try-finally block, but I'm not fond of that syntax, because if you start passing around pooled resources between classes and methods then it's going to get very confusing. It's possible that the main class that uses a resource doesn't even have a reference to the pool. It really becomes quite messy, so a better approach is to create a "smart" pooled object.
Let's say we start with the following simple interface/class:
public interface IFoo : IDisposable
{
void Test();
}
public class Foo : IFoo
{
private static int count = 0;
private int num;
public Foo()
{
num = Interlocked.Increment(ref count);
}
public void Dispose()
{
Console.WriteLine("Goodbye from Foo #{0}", num);
}
public void Test()
{
Console.WriteLine("Hello from Foo #{0}", num);
}
}
Here's our pretend disposable Foo resource which implements IFoo and has some boilerplate code for generating unique identities. What we do is to create another special, pooled object:
public class PooledFoo : IFoo
{
private Foo internalFoo;
private Pool<IFoo> pool;
public PooledFoo(Pool<IFoo> pool)
{
if (pool == null)
throw new ArgumentNullException("pool");
this.pool = pool;
this.internalFoo = new Foo();
}
public void Dispose()
{
if (pool.IsDisposed)
{
internalFoo.Dispose();
}
else
{
pool.Release(this);
}
}
public void Test()
{
internalFoo.Test();
}
}
This just proxies all of the "real" methods to its inner IFoo (we could do this with a Dynamic Proxy library like Castle, but I won't get into that). It also maintains a reference to the Pool that creates it, so that when we Dispose this object, it automatically releases itself back to the pool. Except when the pool has already been disposed - this means we are in "cleanup" mode and in this case it actually cleans up the internal resource instead.
Using the approach above, we get to write code like this:
// Create the pool early
Pool<IFoo> pool = new Pool<IFoo>(PoolSize, p => new PooledFoo(p),
LoadingMode.Lazy, AccessMode.Circular);
// Sometime later on...
using (IFoo foo = pool.Acquire())
{
foo.Test();
}
This is a very good thing to be able to do. It means that the code which uses the IFoo (as opposed to the code which creates it) does not actually need to be aware of the pool. You can even inject IFoo objects using your favourite DI library and the Pool<T> as the provider/factory.
I've put the complete code on PasteBin for your copy-and-pasting enjoyment. There's also a short test program you can use to play around with different loading/access modes and multithreaded conditions, to satisfy yourself that it's thread-safe and not buggy.
Let me know if you have any questions or concerns about any of this.
