What Loom and Go do is to schedule green threads on a bunch of system threads and spawn more system threads when they get blocked doing synchronous system calls.
(And if you want to be pedantic, IIRC the green threads were originally mapped to the Java process, not a thread, because threads were either unavailable or immature in Linux when Java 1 came out; I can’t remember which)
The green threads on the JVM were not the same kind as green threads in Go (and Loom), they would block on IO. Can't speak for Loom, but Go automagically reschedules your green thread when it blocks which allows other threads to run while waiting.
The point is they weren't rescheduled when they would block in the JVM, every process has a main thread.
But I find the cycle - what I called the hype cycle - from internal scheduling to external scheduling and back again, interesting, and I wonder what, if anything, we as an industry can learn from this?
ISTM that Java 1.2 could have improved on green threads instead of moving to os threads. So, is there something we can learn from these two transitions that will help us all make better decisions in the future? The use of OS threads and all the complexity that this has caused has cost the industry hundreds of thousands of hours of developer time. If we can learn some lessons from this then isn’t that a good thing?
1. Threading got much faster and lightweight. This is what Java was initially trying to work around, until it didn't have to any more.
2. The problem moved to handling as many sockets concurrently as possible. Even lightweight system threads are too heavy for scaling linearly with the number of connections (too much context-switching overhead, too much space for stack, etc.)
Green-threading has become a good idea again because we now have a kernel API that is used to multiplex a lot (but not all) I/O systemcalls.
Today Go runtime uses epoll/kqueue to read from a big bunch of sockets, whenever something new happens to any of them. This takes one system thread only.
The API model of epoll/kqueue implies some way to handle concurrency in your user code: this can either be callbacks (or async/await syntactic sugar) or green threads and CSP (channels and so on.) This is why green threading is having a comeback.
(Sorry for implying you did not read the article!)
> Green-threading has become a good idea again because we now have a kernel API that is used to multiplex a lot (but not all) I/O systemcalls.
OK, that makes a lot of sense. I had to read up about how epoll is different from select/poll (that's how long it's been since I worked in C :). Clearly epoll was needed to make green threads efficient, but from what I can see, by the time epoll and friends were widespread, the pthread model was entrenched in Java.
No it’s not the same thing - Java was previously M:1, then became M:M, and now Loom is M:N and also preserves M:M at the same time.
It’s a comment on the hype cycle. As a long time (since 1.0!) java developer, I like that Java has adopted a more efficient threading model. But they are certainly returning to a state where the JVM manages the threads rather than the OS, which certainly is the same as the original release, at least on Linux.
So it's more circumstance and necessity than hype.
Today, green threads are faster. Have os threads become slower? Not that I know of. All that’s really changed is how green threads are implemented.
We went from “green threads bad os threads good” to “os threads bad green threads good”. Sure, the new green threads are better than the old ones, but in some alternate universe maybe we could have got here without os threads (and maybe saved hundreds of thousands of hours of developer time in the process).
To me, what’s interesting is that this appears to be a case of evolutionary “local maxima”. Java went from green threads to os threads because os threads gave better performance for the amount of effort expended. But now we’re moving back to green threads - at much greater effort, 20 years later. I’m sure I’m not the only person that wonders what lessons we can learn from this?
Green threads employed only one kernel thread so couldn't take advantage of multicore, and at the time Java relied on native code so much more than today.
> but in some alternate universe maybe we could have got here without os threads
Maybe :)
> But now we’re moving back to green threads - at much greater effort, 20 years later
These aren't green threads but M:N, and I don't think it's a much greater effort than the effort that would have been required to add it to HotSpot back then.
The first green threads implementation, which was of course Java 1.0, was certainly M:1 - my memory is that we didn't even have pthreads in linux at the time - but the term today doesn't specifically mean a M:1 mapping.
On the other hand, of course JVM1.0 invented "green threads" which was M:1. So yeah this is confusing and I made it more confusing. Sorry. :)
So the point I was trying to make is that Java originally scheduled threads in userspace, then delegated scheduling to the OS, and has now taken it back. In hindsight it seems that delegating this key function to the OS has caused a lot of pain, and I'm interested in the drivers for using OS threads instead of just making the original green thread implementation better.
To be even clearer, I see this a lot in software. We delegate some key function of a product (like database access) to some inefficient and complex machinery (like an ORM) and then spend years working around the problems that decision caused. I've done this myself. I'd like to understand how to stop doing it.
I don't know all the reasons behind abandoning green threads, but some are mentioned here: https://docs.oracle.com/cd/E19455-01/806-3461/6jck06gqe/inde... namely, no multi-core and bad FFI ("interoperate with existing MT applications"). Plus, a little later, the transition to HotSpot left little desire to reimplement userspace threading. There was no pressing need, there was other, higher-priority work needed, and FFI was still an impediment. In other words, making green threads better was neither an important requirement nor a reasonable possibility back then.
Why have we "gone back"? 1. the scale of concurrency required today is higher and resulted in the scaffolding I mention in the article, 2. FFI is no longer an impediment as Java code rarely relies on native code these days.
Nevertheless, Project Loom doesn't switch from kernel to userspace threads, but rather lets you choose which implementation you want for each thread.
We aren’t moving back! Back then it was M:1. We're now moving to M:N/M:M.
You’re making it sound like we’re reverting to the previous design or implementation. Neither is true.
(Also, IIRC Linux didn't even have native OS threads when Java 1.0 came out, so M:1 was literally and the only option available in Linux at the time. My memory is that we had to wait for pthreads before we could have hotspot on Linux, but it's a long time ago and quite tangental to my point).
This is where you're getting confused.
This wasn't an 'intervening decision' because we haven't removed it and with Loom the OS is still able to schedule your threads. The user-space threads are opt-in. That's why it's not a reversion. Because Java was never like this before.
I thought this was weird because IIRC Java 1.0 used green threads in Linux and it was a big deal when they moved to OS threads.
I’ve long believed that the IT world moves in cycles but this is a very clear example of exactly that. Java has gone from green threads to Posix threads and now back to green threads.
I do think it’s awesome (I love goroutines) and Threads in Java have become a but of a nightmare, made a little easier with executors and CompletableFutures. So this further improvement is great news.
But still... call me when the builder pattern is dead.