I’ve been playing with collections and threading and came across the nifty extension methods people have created to ease the use of ReaderWriterLockSlim by allowing the IDisposable pattern.
However, I believe I have come to realize that something in the implementation is a performance killer. I realize that extension methods are not supposed to really impact performance, so I am left assuming that something in the implementation is the cause… the amount of Disposable structs created/collected?
Here’s some test code:
using System;
using System.Collections.Generic;
using System.Threading;
using System.Diagnostics;
namespace LockPlay {
static class RWLSExtension {
struct Disposable : IDisposable {
readonly Action _action;
public Disposable(Action action) {
_action = action;
}
public void Dispose() {
_action();
}
} // end struct
public static IDisposable ReadLock(this ReaderWriterLockSlim rwls) {
rwls.EnterReadLock();
return new Disposable(rwls.ExitReadLock);
}
public static IDisposable UpgradableReadLock(this ReaderWriterLockSlim rwls) {
rwls.EnterUpgradeableReadLock();
return new Disposable(rwls.ExitUpgradeableReadLock);
}
public static IDisposable WriteLock(this ReaderWriterLockSlim rwls) {
rwls.EnterWriteLock();
return new Disposable(rwls.ExitWriteLock);
}
} // end class
class Program {
class MonitorList<T> : List<T>, IList<T> {
object _syncLock = new object();
public MonitorList(IEnumerable<T> collection) : base(collection) { }
T IList<T>.this[int index] {
get {
lock(_syncLock)
return base[index];
}
set {
lock(_syncLock)
base[index] = value;
}
}
} // end class
class RWLSList<T> : List<T>, IList<T> {
ReaderWriterLockSlim _rwls = new ReaderWriterLockSlim();
public RWLSList(IEnumerable<T> collection) : base(collection) { }
T IList<T>.this[int index] {
get {
try {
_rwls.EnterReadLock();
return base[index];
} finally {
_rwls.ExitReadLock();
}
}
set {
try {
_rwls.EnterWriteLock();
base[index] = value;
} finally {
_rwls.ExitWriteLock();
}
}
}
} // end class
class RWLSExtList<T> : List<T>, IList<T> {
ReaderWriterLockSlim _rwls = new ReaderWriterLockSlim();
public RWLSExtList(IEnumerable<T> collection) : base(collection) { }
T IList<T>.this[int index] {
get {
using(_rwls.ReadLock())
return base[index];
}
set {
using(_rwls.WriteLock())
base[index] = value;
}
}
} // end class
static void Main(string[] args) {
const int ITERATIONS = 100;
const int WORK = 10000;
const int WRITE_THREADS = 4;
const int READ_THREADS = WRITE_THREADS * 3;
// create data - first List is for comparison only... not thread safe
int[] copy = new int[WORK];
IList<int>[] l = { new List<int>(copy), new MonitorList<int>(copy), new RWLSList<int>(copy), new RWLSExtList<int>(copy) };
// test each list
Thread[] writeThreads = new Thread[WRITE_THREADS];
Thread[] readThreads = new Thread[READ_THREADS];
foreach(var list in l) {
Stopwatch sw = Stopwatch.StartNew();
for(int k=0; k < ITERATIONS; k++) {
for(int i = 0; i < writeThreads.Length; i++) {
writeThreads[i] = new Thread(p => {
IList<int> il = p as IList<int>;
int c = il.Count;
for(int j = 0; j < c; j++) {
il[j] = j;
}
});
writeThreads[i].Start(list);
}
for(int i = 0; i < readThreads.Length; i++) {
readThreads[i] = new Thread(p => {
IList<int> il = p as IList<int>;
int c = il.Count;
for(int j = 0; j < c; j++) {
int temp = il[j];
}
});
readThreads[i].Start(list);
}
for(int i = 0; i < readThreads.Length; i++)
readThreads[i].Join();
for(int i = 0; i < writeThreads.Length; i++)
writeThreads[i].Join();
};
sw.Stop();
Console.WriteLine("time: {0} class: {1}", sw.Elapsed, list.GetType());
}
Console.WriteLine("DONE");
Console.ReadLine();
}
} // end class
} // end namespace
Here’s a typical result:
time: 00:00:03.0965242 class: System.Collections.Generic.List`1[System.Int32] time: 00:00:11.9194573 class: LockPlay.Program+MonitorList`1[System.Int32] time: 00:00:08.9510258 class: LockPlay.Program+RWLSList`1[System.Int32] time: 00:00:16.9888435 class: LockPlay.Program+RWLSExtList`1[System.Int32] DONE
As you can see, using the extensions actually makes the performance WORSE than just using lock (monitor).
Looks like its the price of instantiating millions of structs and the extra bit of invocations.
I would go as far as to say that the ReaderWriterLockSlim is being misused in this sample, a lock is good enough in this case and the performance edge you get with the ReaderWriterLockSlim is negligible compared to the price of explaining these concepts to junior devs.
You get a huge advantage with reader writer style locks when it takes a non-negligable amount of time to perform reads and writes. The boost will be biggest when you have a predominantly read based system.
Try inserting a Thread.Sleep(1) while the locks are acquired to see how huge a difference it makes.
See this benchmark:
In my benchmarking I do not really notice much of a difference between the two styles, I would feel comfortable using it provided it had some finalizer protection in case people forget to dispose ….