Fast Allocations in Ruby 3.5
railsatscale.com254 points by tekknolagi a day ago
254 points by tekknolagi a day ago
Ruby keeps on getting better. I wouldn't hesitate to start new projects in ruby.
> I’ve been interested in speeding up allocations for quite some time. We know that calling a C function from Ruby incurs some overhead, and that the overhead depends on the type of parameters we pass.
> it seemed quite natural to use the triple-dot forwarding syntax (...).
> Unfortunately I found that using ... was quite expensive
> This lead me to implement an optimization for ... .
That’s some excellent yak shaving. And speaking up … in any language is good news even if allocation is not faster.
Thank you! Wish it had panned out for Class#new, but I don't feel bad about doing it. :)
Can someone explain, is YJIT being abandoned over the new ZJIT? [0]
And if so, will these YJIT features likes Fast Allocations be brought to ZJIT?
It's not being abandoned, we're just shifting focus to evaluate a new style of compiler. YJIT will still get bug fixes and performance improvements.
ZJIT is a method based JIT (the type of compiler traditionally taught in schools) where YJIT is a lazy basic block versioning (LBBV) compiler. We're using what we learned developing and deploying YJIT to build an even better JIT compiler. IOW we're going to fold some of YJIT's techniques in to ZJIT.
> And if so, will these YJIT features likes Fast Allocations be brought to ZJIT?
It may not have been clear from the post, but this fast allocation strategy is actually implemented in the byte code interpreter. You will get a speedup without using any JIT compiler. We've already ported this fast-path to YJIT and are in the midst of implementing it in ZJIT.
But Aaron, what do you actually do here?!
I’m so glad to see your work, and it’s always such a treat to read any of your new posts. Hope to see upcoming ones more often!
Why is a traditional method based JIT better than an LBBV JIT? I thought YJIT is LBBV because it's a better fit for Ruby, whereas traditional method based JIT is more suitable for static languages like Java.
One reason is that we think we can make better use of registers. Since LBBV doesn't "see" all blocks in a particular method all at once, it's much more challenging to optimize register use across basic blocks. We've added type profiling, so ZJIT can "learn" types from the runtime.
Are you piggybacking on LLVM for registerallocation and things like that or building from scracth?
I can imagine combining BBV type branching info with register tracing would add a lot of complicated structures, but I thought the BB's of BBV was more or less analogous to SSA blocks so no middle ground to be found? (consdering if there are many megamorphic sites in the standard library?)
Usual caveat that while Java is mostly static, is has dynamic runtime semantics inherited from Smalltalk and Objective-C, with dynamic class loading, bytecode generation, proxy classes, reflection, hence why the research work on Smalltalk and Strongtalk ended up being so useful for Hotspot.
Yjit is based on Maxime's basic block vesioning PHD work (a JS JIT), the approach both has taken is very dynamic-type-focused, it's a clever way to basically propagate type-info as the code is built. The main benefit is that you get a sane JIT fairly quickly and it should behave well in most dynamic typing scenarios.
They're pivoting (successfully?) to a more traditional way, letting the interpreter first profile the code (to figure out the types) and THEN produce entire methods with heavier optimizations that can do better register allocation.
The BBV approach is sane out of the box but kinda unfamiliar for many compiler writers (problems hiring?) and probably has some performance ceilings without much complexity.
The major question as to what method will win out depends on this question, how "monomorphic" or "polymorphic" is Ruby code in real life?
Monomorphic basically means that only one "real type"(from the compilers point of view) will ever pass a codepath (and thus extra machinery to allow multiple types won't bring much benefit).
>For this reason, we will continue maintaining YJIT for now and Ruby 3.5 will ship with both YJIT and ZJIT. In parallel, we will improve ZJIT until it is on par (features and performance) with YJIT.
I guess YJIT will always be faster in warmup and minimal increase of memory usage. ZJIT being more traditional should bring more speedup than YJIT.
But most of the speedup right now is still coming from rewriting C into Ruby.
> But most of the speedup right now is still coming from rewriting C into Ruby.
Quick glance, this statement seems backwards - shouldn't C always be faster? or maybe i'm misunderstanding how the JIT truly works
C itself is fast; it's calls to C from Ruby that are slow. [1]
Crossing the Ruby -> C boundary means that a JIT compiler cannot optimize the code as much; because it cannot alter or inline the C code methods. Counterintuitively this means that rewriting (certain?) built-in methods in Ruby leads to performance gains when using YJIT. [2]
[1]: https://railsatscale.com/2023-08-29-ruby-outperforms-c/ [2]: https://jpcamara.com/2024/12/01/speeding-up-ruby.html
Unless your JIT can analyse the full code, a transition between byte code and native code is often costly because the JIT won't be able to optimize the full path. Once your JIT generates good enough code, it then becomes faster to avoid that transition even in cases when in isolation native code might still be faster.
EDIT: Note that this isn't an inherent limit. You could write a JIT that could analyze the compiled C code too. It's just that it's much harder to do.
And that is what TruffleRuby did. I had wished there is a subset of Ruby that could be compiled to C. And then all gems should be written in that instead. I remember a few people tried but failed though. Have to dig up the old HN threads again.
Compiling a subset of Ruby to C wouldn't be that hard, but making it compile to C that is fast enough to be worth it is. Not because the Ruby VM is particularly fast, but because the "naive" way of compiling Ruby to C still incurs almost all of the overhead.
E.g. TruffleRuby is fast in part because it will do things like try to avoid method calls for built in types where the standard operations haven't been overridden, but that requires a lot of extra machinery...
So I'm not sure how much compiling to C would help for gems that use C to speed things up.
I think maybe an easier target would be to compile C to a slightly augmented Ruby bytecode. If you control the C compiler you could do things like make C code follow the Ruby calling convention until/unless calling external C code, and avoid a lot of stack overhead.
Not everyone failed, see RubyMotion.
However they decided it was more useful as a commercial product.
The sibling comments mention that C is used in a lot of places in Ruby that incur cross-language overheads, which is true, but it's also just true that in general, even ignoring this overhead, JIT'd functions are going to be faster then their comparable C functions, because 1) they have more profiling information to be able to work from, 2) they have more type information, and (as a consequence of 1&2) 3) they're more likely to be monomorphized, and the compiler is more able to inline specialized variants of them into different chunks of the code. Among other optimizations!
If you give the JIT compiler unlimited time with the code, then maybe. For real large applications, optimized JIT compiled code tends to lag behind AOT optimized C or Rust code, though I guess you could argue that these differences are language / runtime related rather than compiler related.
> ...they have more profiling information to be able to work from... more type information... more likely to be monomorphized, and the compiler is more able to inline specialized variants of them into different chunks of the code.
this is fascinating to me. i always assumed C had everything in the language that was needed for the compiler to use. in other words, the compiler may have a lot to work through, but the pieces are all available. but this makes it sound like JIT'd functions provide more info to the compiler (more pieces to work with). is there another language besides C that does have language features to indicate to the compiler how to make things as performant as possible?
A very simple way to think about is that if an intrinsic is written in C, the JIT can't easily inline it, whereas all ruby code can be inlined. Inlining is the most important optimization that enables other optimizations.
It's not necessarily the fact that C doesn't have enough information, it's just that the JIT can reason about Ruby code better than it can about C code. To the JIT, C code is just some function which does things and the only thing it can do with it is to call it.
On the other hand, a Ruby function's bytecode is available to the jit, so if it sees fit, it can copy paste the function body into the call site and eliminiate the function call overhead. Further, after the inlining, it can apply a lot of further optimizations across what was previously a function boundary.
In theory, you could have a way to "compile" the C intrinsics into the JIT's IR directly and that would also give you similar results.
JITs have runtime information that static compilers do not. Sometimes that's not a huge benefit, but it can often have massive performance implications. For example, a JIT can inline dynamically loaded code into your own code. That sounds unusual, but it's actually ultra-common in practice. For example, this shows up in something as mundane and simple as configurable logging.
The c code in question is most likely interpreter code that is incredibly generic meaning it is very branchy based on data that is only known at runtime, and therefore can’t be optimized at compile time. Jit has the benefit of running the compiler at runtime when the data is known.
C is actually a pretty hard language to compile well. C is a very weakly typed language (e.g. malloc returns a void* that the user manually casts to the type they intended), and exposes raw pointers to the user, which makes analysis for compilers really annoying.
It doesn't sound like YJIT is being abandoned at all. Reading between the lines, it sounds like they want to invest most of their new development right now into a less-experimental architecture that's closer to other JITs and is easier to develop, but that they consider this a somewhat risky endeavor and aren't sure whether this investment is going to pan out in the long run. So they're going to try ZJIT out, but YJIT and the ideas behind it are by no means abandoned. They're just taking a pause to see if a rewrite will make it easier to maintain or produce better results in the long term.
> It’s very rare for code to allocate exactly the same type of object many times in a row, so the class of the instance local variable will change quite frequently.
That’s dangerous thinking because constructors will be a bimodal distribution.
Either a graph of calls or objects will contain a large number of unique objects, layers of alternating objects, or a lot of one type of object. Any map function for instance will tend to return a bunch of the same object. When the median and the mean diverge like this your thinking about perf gets muddy. An inline cache will make bulk allocations in list comprehensions faster. It won’t make creating DAGs faster. One is better than none.
> Any map function for instance will tend to return a bunch of the same object.
Yes, but if it ends up creating any ephemeral objects in the process of determining those returned objects, then the allocation sequence is still not homogeneous. In Ruby, according to the article, even calling a constructor with named arguments allocates, so it's very easy to still end up cycling through allocating different types.
At the same time, the callsite for any given `.new()` invocation will almost always be creating an instance of the exact same class. The target expression is nearly always just a constant name. That makes it a prime candidate for good inline caching at those callsites.
> Yes, but if it ends up creating any ephemeral objects in the process of determining those returned objects, then the allocation sequence is still not homogeneous.
Yes! People might do `map` transformations, but it's very common to do other stuff at the same time. Any other allocations during that transformation would ruin cache hit rate.
> At the same time, the callsite for any given `.new()` invocation will almost always be creating an instance of the exact same class. The target expression is nearly always just a constant name. That makes it a prime candidate for good inline caching at those callsites.
Yes again!
This is why it's imperative that inline caches learn and adapt to the observed behavior. As long as learning is cheap, identifies profitable cases effectively, and backs off for polymorphic and megamorphic scenarios, it's a win.
VM implementer intuition only goes so far, and as the internet is the greatest fuzzer invented, you're definitely going to encounter programs that break your best laid plans.
> This is why it's imperative that inline caches learn and adapt to the observed behavior.
True, but if you only have a single bottleneck cache site for all constructor invocations across the program, the only reasonable thing that callsite can learn is "wow, every single constructed class goes through here".
That's why it makes sense to have a separate cache at every `.new()` location.
Yeah, then you want context-sensitive ICs which are indexed by the callsite. JSC gets some of this by profiling in higher tiers, where inlining might have occurred.
> One is better than none.
Not necessarily. An inline cache is cheap but it's not free, even less so when it also comes with the expense of moving Class#new from C to Ruby. It's probably not worth speeding up the 1% at the expense of the 99%.
> An inline cache will make bulk allocations in list comprehensions faster.
Only if such comprehensions create exactly one type of object, if they create two it's going to slow them down, and if they create zero (just do data extraction) it won't do anything.
> Only if such comprehensions create exactly one type of object,
We just had this conversation maybe a month ago. If it’s 50-50 then you are correct. However if it’s skewed then it depends. I can’t recall what ratio was discovered to be workable, it was more than 50% and less than or equal to 90%.
It seems to me like all languages are converging towards something like WASM. I wonder if in 20 years we will see WASM become the de facto platform that all apps can compile to and all operating systems can run near-natively with only a thin like WASI but more convenient.
This has been an idea since UNCOL was discussed as idea back in 1958.
https://en.wikipedia.org/wiki/UNCOL
There are countless bytecode based platforms since 1958, including all famous Xerox PARC systems (the CPUs were microcoded and loaded the related translation code on boot), yet WASM is doing it first keeps being brought up.
Do you know what I call WASI containers running on a Kubernetes cluster, or serverless cloud vendors?
Application Server, https://en.wikipedia.org/wiki/Application_server
Multi-tenancy on the JVM makes me shudder, though. The general point you're making, thumbs up to that.
Wasn't this the idea of the JVM?
Java bytecode was originally never intended to be used with anything other than Java - unlike WASM it's very much designed to describe programs using virtual dispatch and automatic memory management. Sun eventually added stuff like invokedynamic to make it easier to implement dynamic languages (at the time, stuff like Ruby and Python), but it was always a bit of round peg in square hole.
By comparison, WASM is really more like traditional assembly, only running inside a sandbox.
Just like CLR bytecode, IBM i TIMI bytecode and many others since 1958.
For some reason when people advocate for WASM outside of the browser, they only remember of the JVM.
And of course the ill-fated Parrot VM associated with the Perl 6 project.
I think that was more of a language-oriented effort rather than runtime/abi oriented effort.
Parrot was intended to be a universal VM. It wasn’t just for Perl.
Sure, I just think that's a very odd way to characterize the project. Basically anything can be universal vm if you put enough effort to reimplementing the languages. Much of what sets Parrot aside is its support for frontend tooling.
“The Parrot VM aims to be a universal virtual machine for dynamic languages…”
That’s how the people working on the project characterized it.
I certainly think the humor in parrot/rakudo (and why they come up today still) is how little of their own self image the proponents could perceive. The absolute irony of thinking that perl's strength was due to familiarity with text-manipulation rather than the cultural mass....
It’s not a bad idea. Lot of the same people who worked on JVM were around while the asm - wasm ideas emerged
Like predicted in 2014 here: https://www.destroyallsoftware.com/talks/the-birth-and-death...
did it means more speeds to all rails/active records collections?
I know I may be jumping the gun a little here but I wonder what percentage speedup could we expect on typical rails applications. Especially with Active Record.
so far no diff here (https://speed.yjit.org/). But the build is from May 14 so maybe it will show up in new build?