You asked:
- What's the difference with using
concurrent + barrier vs just using a serialQueue for setting in this case?
The former allows concurrent reads and the latter does not. But both are standard patterns for thread-safety.
FWIW, this “concurrent + barrier” approach is known as the “reader-writer” pattern. It features two key behaviors, namely that reads can happen concurrently (because it is a concurrent queue) and also that the caller does not have to wait for writes (because we invoke it with async with barrier).
- I did a testing on playground, I wrapped the get and set with 1000 time for loop, it turns out that the behavior for both serial and concurrent queue is almost the same.
First, Playgrounds are not an ideal environment for testing subtle performance differences like this. Test in an app with a optimized, “release” build. (This also gives you the chance to debug with tools like the “Thread Sanitizer”.)
Second, you will only see very modest performance differences (probably only observable if doing millions of iterations, not thousands). But the overall behavior of these two patterns should be similar, as they are doing the same thing, namely synchronizing your access to this dictionary.
Third, your two for loops invoke this MutableDictionary from a single thread. If you are testing a thread-safe dictionary, I would suggest actually calling it from multiple threads. There is no point in introducing the thread-safety overhead if you are only accessing it from a single thread. E.g., this tests multithreaded behavior:
let dict = MutableDictionary()
DispatchQueue.concurrentPerform(iterations: 1_000) { i in
let result = dict.object(for: "\(i)")
print(result) // this will always be `nil`, though, because your dictionary has no values yet
}
DispatchQueue.concurrentPerform(iterations: 1_000) { i in
dict.set("a", for: "\(i)")
}
DispatchQueue.concurrentPerform is a parallel for loop.
- Why the setting always raise an error?
This is introduced by Playgrounds, probably its inability to handle the thread explosion of the unbridled dispatching of 1,000 asynchronous tasks. Try it in an actual app, not a playground, and the worker thread exhaustion will not result in this crash.
That having been said, you should probably avoid scenarios that permit thread explosion at all.
For example, if you really want to have a thread update 1,000 values in the dictionary, all from a single thread, you might provide a method that allows you to update multiple values with a single synchronization call. E.g.,
class MutableDictionary<Key: Hashable, Value> {
private var dictionary: [Key: Value] = [:]
private let queue = DispatchQueue(label: "queue", attributes: .concurrent)
func object(for key: Key) -> Value? {
queue.sync {
dictionary[key]
}
}
func set(_ object: Value?, for key: Key) {
queue.async(flags: .barrier) {
self.dictionary[key] = object
}
}
func synchronized(block: @escaping (inout [Key: Value]) -> Void) {
queue.async(flags: .barrier) { [self] in
block(&dictionary)
}
}
}
let dictionary = MutableDictionary<Int, String>()
dictionary.synchronized {
for i in 0 ... 1_000 {
$0[i] = "a"
}
}
That updates the thousand dictionary values in a single synchronization. This simultaneously eliminates unnecessary synchronizations and solves the thread explosion issue.
- What does concurrent queue do better in this case (for both get and set) compared to a serial queue?
Theoretically, the concurrent queue, “reader-writer” pattern, can offer better performance if you have an app that may be doing many concurrent reads simultaneously. Otherwise the overhead of the concurrent queue is unlikely not necessary/beneficial.
In practice, I have yet to encountered real-world scenarios where the concurrent queue was observably faster.
- In iOS, when should I choose serial queue over concurrent queue? And vice-versa?
I would benchmark it for your particular use-case and see if the concurrent queue yields observable performance benefits. If so, the overhead of the concurrent queue might be worth it. Otherwise, you might want to stick with the simple serial GCD queue.
Generally, you will not see an observable difference. And where I have, I found that other approaches (e.g. NSLock or os_unfair_lock) were even faster. But if you are looking for a simple and robust synchronization mechanism, a serial dispatch queue is a good solution.
A few other observations:
If creating a dictionary, I would be inclined to avoid the use of Any. We have a strongly-typed language and it is a shame to lose these benefits. E.g., you might make it a generic as shown in the example above with Key and Value.
The wrapped dictionary should be private. It is imprudent to expose the underlying dictionary which would undermine the thread-safe interface.
The synchronization mechanism (e.g., the lock or GCD queue or whatever) should be private, too.
