tabs ↹ over ␣ ␣ ␣ spaces

by Jiří {x2} Činčura

How slow the “lock” statement actually is?

26 Oct 2014 C#, Multithreading/Parallelism/Asynchronous/Concurrency

When I’m teaching my “parallel-threading-asynchronous-locking-synchronization” course I discuss the lock statement a lot. Not only because it’s probably the most widely used “lock”, but also because it’s pretty easy to fall into a trap if you don’t know what’s going on behind the scenes and around you.

Anyway. I often show how the lock is translated into try-finally block and that it does not come for free. And among other things I’m trying to get into the brains of the people I’m teaching is that you should spent as less time as possible inside lock (of course, with some exceptions) - less contention equals less waiting hence more resources used for real work.

I kind of know the finally block isn’t cheap. But how much? Does it really make sense to try to avoid it? Last night, being in the hotel, I decided to test it. As with every test it’s one particular scenario and the mileage might vary.

I created a simple test code.

class Program
{
	static object SyncRoot = new object();
	static int Dummy = 0;

	static void Main(string[] args)
	{
		const long TestsCount = 400000000;
		Stopwatch sw;

		TimeSpan lockElapsed;
		TimeSpan monitorElapsed;

		LockTest();
		MonitorTest();

		sw = Stopwatch.StartNew();
		for (long i = 0; i < TestsCount; i++)
		{
			LockTest();
		}
		lockElapsed = sw.Elapsed;

		sw = Stopwatch.StartNew();
		for (long i = 0; i < TestsCount; i++)
		{
			MonitorTest();
		}
		monitorElapsed = sw.Elapsed;

		Console.WriteLine("Lock    : {0}", lockElapsed);
		Console.WriteLine("Monitor : {0}", monitorElapsed);
		Console.WriteLine("%       : {0}", (double)lockElapsed.Ticks / (double)monitorElapsed.Ticks);

		Console.WriteLine(Dummy);
	}

	static void LockTest()
	{
		lock (SyncRoot)
		{
			Dummy++;
		}
	}

	static void MonitorTest()
	{
		Monitor.Enter(SyncRoot);
		{
			Dummy++;
		}
		Monitor.Exit(SyncRoot);
	}
}

Really nothing special (I hope I haven’t overlooked something important). I get some stopwatch to measure the time, warm up the methods and start running in a loop. I did few test runs (discarding numbers that were very off) without debugger attached and (of course) with optimizations turned on. I also tried how the 32bit vs 64bit (“Prefer 32-bit” checkbox) affects the result.

So what’s the numbers? As expected using the Monitor.Enter and Monitor.Exit is faster than using lock statement. I saw numbers between 15 % and 17 %. With “Prefer 32-bit” turned on. Surprisingly with "Prefer 32-bit" turned off (and on a 64bit OS) the difference was only about between <1 % and 3 %. Pretty interesting.

There’s one catch, though. The lock statement translation is not what I wrote in MonitorTest method (among the try-finally). It actually uses the Monitor.Enter overload with the ref bool lockTaken parameter. So the method would rather be.

static void MonitorTest()
{
	var lockTaken = false;
	Monitor.Enter(SyncRoot, ref lockTaken);
	{
		Dummy++;
	}
	if (lockTaken)
		Monitor.Exit(SyncRoot);
}

Is there a change in results now? Yes, a bit. Again the 32bit first. Now it was between about 17 % and 18 %. But the 64bit was between 0 % (few times the lock was even faster, though it was on a measurement error boundary) and 2 %.

Now you might think that few percent difference on a μs operation is not important. Sure. But also take into account that when you start writing some locking or synchronization you want it to be fast. To utilize all resources you have available for getting the result. That means (not only) getting in and out the lock as fast as you can.

As usual the decision is on you.

Part 2 of this story.