Why is the explicit management of threads a bad thing?

In a previous question, I made a bit of a faux pas. You see, I'd been reading about threads and had got the impression that they were the tastiest things since kiwi jello.

Imagine my confusion then, when I read stuff like this:

[T]hreads are A Very Bad Thing. Or, at least, explicit management of threads is a bad thing

and

Updating the UI across threads is usually a sign that you are abusing threads.

Since I kill a puppy every time something confuses me, consider this your chance get your karma back in the black...

How should I be using thread?

Solution 1:

Enthusiam for learning about threading is great; don't get me wrong. Enthusiasm for using lots of threads, by contrast, is symptomatic of what I call Thread Happiness Disease.

Developers who have just learned about the power of threads start asking questions like "how many threads can I possible create in one program?" This is rather like an English major asking "how many words can I use in a sentence?" Typical advice for writers is to keep your sentences short and to the point, rather than trying to cram as many words and ideas into one sentence as possible. Threads are the same way; the right question is not "how many can I get away with creating?" but rather "how can I write this program so that the number of threads is the minimum necessary to get the job done?"

Threads solve a lot of problems, it's true, but they also introduce huge problems:

• Performance analysis of multi-threaded programs is often extremely difficult and deeply counterintuitive. I've seen real-world examples in heavily multi-threaded programs in which making a function faster without slowing down any other function or using more memory makes the total throughput of the system smaller. Why? Because threads are often like streets downtown. Imagine taking every street and magically making it shorter without re-timing the traffic lights. Would traffic jams get better, or worse? Writing faster functions in multi-threaded programs drives the processors towards congestion faster.

What you want is for threads to be like interstate highways: no traffic lights, highly parallel, intersecting at a small number of very well-defined, carefully engineered points. That is very hard to do. Most heavily multi-threaded programs are more like dense urban cores with stoplights everywhere.

• Writing your own custom management of threads is insanely difficult to get right. The reason is because when you are writing a regular single-threaded program in a well-designed program, the amount of "global state" you have to reason about is typically small. Ideally you write objects that have well-defined boundaries, and that do not care about the control flow that invokes their members. You want to invoke an object in a loop, or a switch, or whatever, you go right ahead.

Multi-threaded programs with custom thread management require global understanding of everything that a thread is going to do that could possibly affect data that is visible from another thread. You pretty much have to have the entire program in your head, and understand all the possible ways that two threads could be interacting in order to get it right and prevent deadlocks or data corruption. That is a large cost to pay, and highly prone to bugs.

• Essentially, threads make your methods lie. Let me give you an example. Suppose you have:

  if (!queue.IsEmpty) queue.RemoveWorkItem().Execute();

Is that code correct? If it is single threaded, probably. If it is multi-threaded, what is stopping another thread from removing the last remaining item after the call to IsEmpty is executed? Nothing, that's what. This code, which locally looks just fine, is a bomb waiting to go off in a multi-threaded program. Basically that code is actually:

 if (queue.WasNotEmptyAtSomePointInThePast) ...

which obviously is pretty useless.

So suppose you decide to fix the problem by locking the queue. Is this right?

lock(queue) {if (!queue.IsEmpty) queue.RemoveWorkItem().Execute(); }

That's not right either, necessarily. Suppose the execution causes code to run which waits on a resource currently locked by another thread, but that thread is waiting on the lock for queue - what happens? Both threads wait forever. Putting a lock around a hunk of code requires you to know everything that code could possibly do with any shared resource, so that you can work out whether there will be any deadlocks. Again, that is an extremely heavy burden to put on someone writing what ought to be very simple code. (The right thing to do here is probably to extract the work item in the lock and then execute it outside the lock. But... what if the items are in a queue because they have to be executed in a particular order? Now that code is wrong too because other threads can then execute later jobs first.)

• It gets worse. The C# language spec guarantees that a single-threaded program will have observable behaviour that is exactly as the program is specified. That is, if you have something like "if (M(ref x)) b = 10;" then you know that the code generated will behave as though x is accessed by M before b is written. Now, the compiler, jitter and CPU are all free to optimize that. If one of them can determine that M is going to be true and if we know that on this thread, the value of b is not read after the call to M, then b can be assigned before x is accessed. All that is guaranteed is that the single-threaded program seems to work like it was written.

Multi-threaded programs do not make that guarantee. If you are examining b and x on a different thread while this one is running then you can see b change before x is accessed, if that optimization is performed. Reads and writes can logically be moved forwards and backwards in time with respect to each other in single threaded programs, and those moves can be observed in multi-threaded programs.

This means that in order to write multi-threaded programs where there is a dependency in the logic on things being observed to happen in the same order as the code is actually written, you have to have a detailed understanding of the "memory model" of the language and the runtime. You have to know precisely what guarantees are made about how accesses can move around in time. And you cannot simply test on your x86 box and hope for the best; the x86 chips have pretty conservative optimizations compared to some other chips out there.

That's just a brief overview of just a few of the problems you run into when writing your own multithreaded logic. There are plenty more. So, some advice:

• Do learn about threading.
• Do not attempt to write your own thread management in production code.
• Use higher-level libraries written by experts to solve problems with threads. If you have a bunch of work that needs to be done in the background and want to farm it out to worker threads, use a thread pool rather than writing your own thread creation logic. If you have a problem that is amenable to solution by multiple processors at once, use the Task Parallel Library. If you want to lazily initialize a resource, use the lazy initialization class rather than trying to write lock free code yourself.
• Avoid shared state.
• If you can't avoid shared state, share immutable state.
• If you have to share mutable state, prefer using locks to lock-free techniques.

Solution 2:

Explicit management of threads is not intrinsically a bad thing, but it's frought with dangers and shouldn't be done unless absolutely necessary.

Saying threads are absolutely a good thing would be like saying a propeller is absolutely a good thing: propellers work great on airplanes (when jet engines aren't a better alternative), but wouldn't be a good idea on a car.