Why lock statements don't scale
- by Alex.Davies
We are going to have to stop using lock statements one day. Just like we had to stop using goto statements. The problem is similar, they're pretty easy to follow in small programs, but code with locks isn't composable. That means that small pieces of program that work in isolation can't necessarily be put together and work together. Of course actors scale fine :) Why lock statements don't scale as software gets bigger Deadlocks. You have a program with lots of threads picking up lots of locks. You already know that if two of your threads both try to pick up a lock that the other already has, they will deadlock. Your program will come to a grinding halt, and there will be fire and brimstone. "Easy!" you say, "Just make sure all the threads pick up the locks in the same order." Yes, that works. But you've broken composability. Now, to add a new lock to your code, you have to consider all the other locks already in your code and check that they are taken in the right order. Algorithm buffs will have noticed this approach means it takes quadratic time to write a program. That's bad. Why lock statements don't scale as hardware gets bigger Memory bus contention There's another headache, one that most programmers don't usually need to think about, but is going to bite us in a big way in a few years. Locking needs exclusive use of the entire system's memory bus while taking out the lock. That's not too bad for a single or dual-core system, but already for quad-core systems it's a pretty large overhead. Have a look at this blog about the .NET 4 ThreadPool for some numbers and a weird analogy (see the author's comment). Not too bad yet, but I'm scared my 1000 core machine of the future is going to go slower than my machine today! I don't know the answer to this problem yet. Maybe some kind of per-core work queue system with hierarchical work stealing. Definitely hardware support. But what I do know is that using locks specifically prevents any solution to this. We should be abstracting our code away from the details of locks as soon as possible, so we can swap in whatever solution arrives when it does. NAct uses locks at the moment. But my advice is that you code using actors (which do scale well as software gets bigger). And when there's a better way of implementing actors that'll scale well as hardware gets bigger, only NAct needs to work out how to use it, and your program will go fast on it's own.
