Could you elaborate on what you consider the interface difference to be?
A thread per connection isn't wrong, though. You're begging the question by assuming that green threads are faster than native threads. I'm specifically arguing against that.
The next point might be if it's harder to get the same performance characteristics with native threads on different platforms.
As I view them, green threads are a useful abstraction just like objects or functions. Do you need something computed, not necessarily now, but at some point? Just make a green thread to compute it and check on the result later. That's the platonic ideal of green threads, anyway: very few languages have gotten to that point (Erlang, maybe Go, on a good day; C# has this with its async functionality but none of it happens in parallel).
But you can't. Green threads have scalability limitations too, and their performance is not predictable (except possibly by the creators of the runtime) in any language or on any processor that does automatic preemption. That's like saying you use i64 instead of i32 because now you can do calculations without worrying about overflow. It may be true for some classes of problem, but it's hardly a different programming paradigm like you're making it out to be. They're definitely not a different abstraction from threads (which are themselves an abstraction!).
while (!ready)
assert(!syscall(SYS_futex, &ready, FUTEX_WAIT, 0, NULL, NULL, 0));
For completeness, I also tested a similar Go program on Go 1.4 (the version available from Debian on the VM); it actually got up to 3,150,000 before OOM, and took 9 seconds to do 2 million - 4.5 microseconds and 2.7KB per thread.
In other words, Linux is about an order of magnitude slower at managing a lrge number of threads. That looks pretty bad, but on the other hand, it's not that much in absolute terms! I'm pretty sure most server programs don't need more than 250,000 simultaneous connections (or can afford to spend more than 8GB of RAM on them) and don't mind spending an extra 20 microseconds to initiate a connection, so if operating systems other than Linux aren't a concern, they could be written to create a thread per connection without too much trouble. It's not going to give you the absolute maximum performance (meaning it's not appropriate for a decent class of program - then again, I suspect Go isn't either), but it's not terrible either.
I'd like to see it improve. I wouldn't be surprised if there is (still) some low hanging fruit; do kernel developers actually care about this use case?
(And yes, I know this doesn't really test performance of the scheduler during sustained operation. That's its own can of worms.)
Yeah, this matches our results when we did similar tests. It's definitely faster to use green threads if you're just spawning and shutting down the threads, but if you're actually doing I/O work on those threads, the overhead quickly drops down.
It's not the fastest way to do I/O, but Go's approach isn't either. The fastest way to do I/O is to forego the thread management syscalls and the stack, like nginx does.
Though Go's is faster.
>forego the thread management syscalls and the stack, like nginx does.
The problem is nginx uses state machines, which aren't a general solution. Stackless coroutines allow you to write functions that the compiler (or a clever library) transforms into state machines.
Okay, so the test you did doesn't actually reflect the use case in practice. Can I expect to reach 200,000 threads if the threads are not all created at exactly the same moment? What if (God forbid) they're doing memory allocation? And if it does work out, will everything be handled efficiently?
...in fact, neither did removing the setstacksize call and having default 8MB stacks. I guess this makes sense: of course the extra VM space reserved for the stacks doesn't require actual RAM to back it; there is some page table overhead, but I guess it's not enough to make a significant difference at this number of allocations. Of course, on 32-bit architectures this would quickly exhaust the address space.
If increasing max_map_count hadn't worked, it would still be possible to allocate stacks on the fly - but you would get a bunch of them in one mmap() call, and therefore in one VM mapping, and dole them out in userland. However, in this case guard pages wouldn't separate different threads' stacks, you would have to generate code that manually checks the stack pointer to avoid security issues from stack overflows, rather than relying on crashing by hitting the guard page. Rust actually already does this, mostly unnecessarily; I'm not sure what Go is doing these days but I think it does too. Anyway, given that the above result I suspect this won't be an actual issue, at least until the number of threads goes up by an order of magnitude or something.
Coroutines have three awesome advantages here:
Really fast switches (it's basically a function call — no need to trap into the kernel)
Scheduling in response to IO events (a coroutine can yield when making a "blocking" IO call and resume when it completes). You can write asynchronous code that looks synchronous!
Better cache behavior (hugely important on modern processors). You can allocate all your coroutine structures together and don't have to suffer all the cache evictions of a context switch every time a coroutine yields.
It would be interesting to see this benchmark repeated with a latency requirement.
You can also write "actual" synchronous code which does the same thing in the kernel. :P And there are several CPU schedulers to choose from to fine-tune when things get woken up.
I agree it would be interesting to test latency; actually, the numbers would be much more useful than my four seconds. After all, if thread creation throughput or latency becomes an issue, you can keep a thread pool around, while still retaining the advantages of one connection per thread.
You can. But no matter how good your kernel scheduler is, that trip through it is going to cost you. Why? Because it's going to wreck your cache and TLB entries. I think the TLB is a big part of the story, given it's small, specific to the process, and particularly expensive to miss.
http://www.cs.cmu.edu/~chensm/Big_Data_reading_group/papers/...
Anyway, I'm sure you know this already. :) Out of curiosity, did you get a chance to take CS169 before you left Brown?
And I think the kernel already does pool/recycle stacks, but I guess not 200,000 of them given your results. An explicit thread pool would certainly work too.
But yeah, certainly userland context switches are faster; it's just that your comment seemed to postulate "scheduling in response to IO events" as a benefit of coroutines over threads, as if the kernel had no way to distinguish between threads waiting for IO and threads needing to be woken up. I presumably misinterpreted it.
Interesting paper. It seems like it could be a useful step toward my ideal imagined environment where the distinction between coroutines and threads would be meaningless - because scheduling would be a job shared between userland and the kernel, so userland could do fast thread switches, batch system calls, etc., while the kernel would still give them PIDs and let them take signals or be ptraced, and native tools (debuggers) would treat them as threads.
Though the paper is from 2010; do you know if anyone has tried to implement anything along similar lines in production?
And no, I didn't take that course.
From what I can tell, M:N threading has been thoroughly abandoned by Linux, but it might be worth revisiting given the horrendous I/O scaling of Linux kernel threads. I would imagine scheduling user threads cooperatively simplifies things a lot.
Google has a very interesting talk about cooperative native thread scheduling, but they haven't upstreamed their code:
https://www.youtube.com/watch?v=KXuZi9aeGTw
And here's what linux devs are actually doing (it's underwhelming, something like 50% of what the hardware is capable of):
https://lwn.net/Articles/629155/
In the meantime, people who actually need the performance (HPC and even some enterprise servers) have to bypass the kernel entirely. It looks like that's going to be possible even in virtualized environments, thanks to hardware support:
http://people.inf.ethz.ch/troscoe/pubs/peter-arrakis-osdi14....
I think that mentality, which I supported, may be what ruined green threads for Rust. Perhaps the enemy was the perfect of the good. On the other hand, maybe not having to worry about differences is the better benefit. I would rather be slower-but-acceptably-fast than have to think about 1:1 vs M:N threading in the case that they are not interchangeable.
On the third hand, it could be that Rust is too close to the metal for green threads to be of any significant benefit. If that's the case, Rust ought to trounce all the M:N implementations, subject to reasonable limits on spawn rate, memory pressure and inter-thread communication. Does that sound right?
Incidentally, I've never seen a cooperative threading implementation that didn't have a yield call of some kind available, and you can usually jam any of them up by writing an infinite computation-only loop, so even in the langauges where the scheduler is automatically invoked, correct 1:1 code can deadlock in the M:N scenario. State machines can too, but a trivial example, translated, might read something like the following:
for(;;);
exit(EXIT_SUCCESS);
It's not meant to be a performance boost, but a way of writing code where "might this crash?" is a question you're never bothered to ask. It's wonderfully freeing.
Not true of native threads. If a goroutine were replaced with a native thread, it would swiftly crash on GC.
Is that actually true? Yes, you won't get a partially written int, but you could still run into consistency issues if you don't lock. For example, if you need thread A to update both x and y to have a consistent state, and B is reading both x and y, A might update x, then the scheduler would switch to B, and be might read the new x and the old y. If the update A was making needed to update both x and y for the system to be in a consistent state, you have a problem.
If you use the int as an offset into raw memory, then sure, that will. But if you use it as an offset into a list, then it won't. At most it will cause an exception to be generated. That will cause the reader thread to die, because it was the code that caused the exception, but it still won't crash.
So I don't think it frees you from thinking about concurrency at all.