Alternatives to using Thread.Sleep for waiting

Question

Firstly I am not asking the same question as C# - Alternative to Thread.Sleep?, or Alternative to Thread.Sleep in C#?. I don't think I am using it incorrectly and need a genuine alternative for specific situations.

During a code analysis run I saw a surprising violation coming up:

Usage of Thread.Sleep() is a sign of flawed design.

This violation leads to Peter Richie's article on why exactly this constitutes bad design.

We all know thread creation is expensive and blocking in threads means contention on the pool. We also know that each thread will allocate a meg of memory so it should have a short lifespan, blocking on the UI is evil, using sleep for timing is unreliable etc etc etc. Which leads me to my point, if you really need to perform a sleep, what should you be using if not Thread.Sleep?

Peter goes on to mention that a zero sleep is the only correct use of Thread.Sleep effectively giving up the thread's timeslice and allowing other threads to process. And then even more scary is that this is only a because of limitations on unmanaged threads and if re-implemented in the CLR will create side effects of using Thread.Sleep in your applications. All the points on common bad usage are, in fact, good examples of bad usage.

I have the following situations in production code that uses Thread.Sleep quite successfully:

• Waiting for a file lock to be relinquished by the operating system (catch file lock problems, wait for a second, try again, give up after a while).
• Killing a process and waiting for it not to show up in the process list (kill it, check it's not running, wait for a second, check it not still running, force it closed).
• Waiting for copy buffers to flush (check a file size, try to access it, wait, check if size has changed).

Without using Thread.Sleep in situations like these, what other options do I have? Tight loops tend to make things worse and I don't believe this makes it's usage a "design flaw" especially since nothing is on the UI and only in background threads. It is just the nature of software to wait for other things in a multi-threaded environment with external factors affecting your code, sometimes you need to wait...

Answer 1

The WaitHandle type and derived types provide an event-driven mechanism for waiting that ties into the operating system. For example, when you have a Task<T> task and you wait on the result by accessing task.Result, the internal implementation isn't polling with Thread.Sleep calls in between. It's using a WaitHandle-derived type to do waiting and synchronization.

Sometimes a polling-based approach is necessary, as in some of the examples you gave in your bullet list, but often you can use an event-driven approach instead. It's not that Thread.Sleep is always bad - it's just that it is very often misused.

It is just the nature of software to wait for other things in a multi-threaded environment with external factors affecting your code, sometimes you need to wait...

To wait is fine. To wait with polling is often not (*). If there is any way you can use an event-driven wait, you should typically strive to use that.

I don't have a very good feel for exactly what it is that you're asking, so I won't elaborate beyond this. If you leave a comment I can expand my answer.

---

(*) The theoretical reason waiting with polling is bad is as follows:

Suppose I have code that looks like this:

//START
Begin();
while (!Done())
    Thread.Sleep(D);
//STOP

Begin() starts some operation. Done() returning true means the operation has finished. Suppose this will happen after approximately T time. Then:

• The thread wakes up and checks the condition (calls Done()) T/D times
• The duration from START to STOP includes an expected D/2 purely because of the Thread.Sleep

What value of D should you choose? As you increase D, the expected duration form START to STOP increases linearly. As you decrease D, the (bound on the) number of iterations increases as 1/D. Both of these are bad, and finding the right D is problematic.

Now compare this to an event-driven wait:

//START
Begin();
WaitDone();
//STOP

Theoretically speaking, as long as WaitDone() somehow magically waits until the operation has finished but no longer, both of the problems identified in the waiting with polling case have disappeared: this thread waits for exactly the right amount of time - no more, no less!

To reiterate the point I started with: in .NET, the WaitHandle class and derived types are what facilitate this approach.

Answer 2

Well, you said most of it. Quoting "We all know thread creation is expensive and blocking in threads means contention on the pool", so you even understand what using a pool of threads is about.

You also understand that blocking the UI thread is bad.

Looking at the pool of threads model again: you have a pool of threads, maybe one per processor, and then pass tasks to them. What sense would it make for one of those threads to block? If it does not have work to perform now, then it should simply proceed to a different task.

So, going directly to your question "Which leads me to my point, if you really need to perform a sleep, what should you be using if not Thread.Sleep?", in a modern well designed program, you would never need to do it, you would simply schedule a task for latter.

You should think of the threads in a pool, much as the processors in a system, as resources, which should be released to others when not needed.

Going to your examples, you're a bit too involved in the imperative programming paradigm.

• You don't need to wait for a process to disappear... I don't imagine why you need this, but if have to wait, it's because you have work to perform after sometime, the "continuation" of your function. You should set up a timer for this "continuation".
• The file examples should have other mechanisms for it, if they do not have.... that would be good OS design. For instance, the only safe way to wait for buffers to flush would be an OS primitive, on the likes of fsync.
• If there is someone writing to a file and then another reading from the file, then a synchronization mechanism is needed, not a timed wait (unless the file is append-only, in which case the file itself is the synchronization mechanism).

Waiting on a synchronization mechanism is not "bad".