How I Made Ruby Faster Than Ruby
noteflakes.com82 points by ciconia 3 days ago
82 points by ciconia 3 days ago
Feels like a misleading headline. The author created another templating language alternative to ERB, found that it was slower than ERB, then optimise it until it was roundabout as fast as ERB given a relatively simple template.
The author then appears to draw the wrong conclusion:
> What I find most interesting about the changes I’ve made to code generation in P2, is that the currently compiled code is more than twice as fast as it was when P2 first came out, which just goes to show than in fact Ruby is not slow, it is actually quite fast, you just need to know how to write fast code! (And I guess this is true for any programming language.)
I love Ruby, but it is still a slow language on most benchmarks. That's ok. For most webapps, the bottleneck is not execution-time performance, it's the speed and joy of writing the code. Functionality that never got built because it was too annoying to build is infinitely slow. But there's no point in pretending Ruby, compared to, say, Rust, isn't a slow-as-molasses execution environment. It is. It's ok. It's optimised for developer happiness, not speed.
And yes, even so, you can write slow Ruby code and fast Ruby code. Again, which one makes sense is contextual. But it doesn't make the point that "Ruby isn't slow."
Hi, author here. Taking your argument to its logical conclusion we can say that it doesn't matter if your Ruby code is slower or faster because it's Ruby, we know it's "slow-as-molasses", we only care about developer happiness, and anyways we're I/O-bound, so it doesn't really matter how our code performs...
But in my experience it does. I've built platforms in Ruby that handle north of 1K reqs/sec with bursts of 10K reqs/sec on moderate hardware, without needing to setup a whole cluster of machines that crunch on poorly-performing code.
From my experience, getting the average execution time per render from say 0.1ms to 0.01ms, and especially reducing allocations and GC pressure has a big effect on 99% percentiles, and consequently on the cost of compute.
Saying because we use Ruby we don't care if it's slow or not is in a way dismissing it as a viable platform for writing reliable software (because performance is part of reliability).
To me, you can use Ruby to create systems that have very good performance characteristics, and still benefit from developer happiness. The two are not contradictory.
> Taking your argument to its logical conclusion we can say that it doesn't matter if your Ruby code is slower or faster because it's Ruby, we know it's "slow-as-molasses", we only care about developer happiness, and anyways we're I/O-bound, so it doesn't really matter how our code performs...
Not OP, but to a point I think this is pretty much true...
We currently have decent performance so it works out well for most use cases, but if Ruby were to be slower, we could probably cover that issue with infra, caching or other means. As we already do in many cases.
It would be a pain point, but in comparison increasing developer happiness or the whole product dev experience is IMHO a lot harder. Perfs would need to be abysmally bad to change that balance.
I’m currently looking at some slow python scripts and libraries that take about 30% of the total build time to generate 20 linker scripts for a project that builds 1,300 C files. Every dev, every build, every CI run is waiting on slow python. So the devs that wrote the tool are happy, but everyone else is slowing dying waiting around for this thing to finish. Relevant [0]
> Every dev, every build, every CI run is waiting on slow python.
Parralelize the build, buy more resource for the CI. It might cost more but it will be "saving lifes" after all, right ?
The question is usually whether those scripts would have existed if it wasn't for Python. I assume if it was trivial to rewrite them you'd do it in 2 hours and go on with your life.
Besides the python scripts, everything is parallelized and is CPU bound on as many cores as it can be given. Because of licensing throwing more CI at it isn’t an option. This is an open source project, so there’s not really money for buy moar bigger.
The tools possibly wouldn’t exist, but there are options now that provide better ergonomics and are not slow.
I think that's the position most "slow" scripts are: They could have been written faster in the first place, but they weren't intended to be kept for so long and/or that wasn't a priority at all in that moment. And now that people are looking at them again they could be fixed, but there's usually still not enough merits to do so.
I assume something will done at some point, but IMHO it's one of these very nice problems to have, as there is a working version that will only be replaced by something better, and only if it's worth it.
Where's the second part of the story: where someone else profiled and discovered 19 of the linker scripts were generated really fast, and that re-working generation of the slowest script only took 15 minutes.
I have been profiling and it’s really just python being slow.The first ⅓ of the run of 12 cores is all python. Fortunately these are mostly independent so there is currently no blocking, but there eventually will be.
Yep, I agree with this.
Jean Boussier wrote this execellent examination of CPU bound Rails application and why making use of multi-processes is probably a good idea.
Even if you're not using CPU bound it's still daft to leave performance on the table you don't need to.
For the most part if something is a bit slower than it needs to be it still makes more sense to take the obvious bottle necks out before you rewrite the whole system in another language. Especially with YJIT and forcoming ZJIT availible.
1. https://byroot.github.io/ruby/performance/2025/01/23/the-myt...
Love all the work Jean Boussier does for the ecosystem.
I would add to this commentary that there are a number of things you can do in Rails to speed it up. For instance, ActionController::Metal
Ruby is slow, but it is however faster than Python last I checked, with like for like code.
Python gives you other interesting ways of going fast (a lot of high performance plugins for numerics, Cython, and so on), while Ruby is a higher level more expressive language IMO, so you have more ways to shoot yourself in the foot more powerfully.
Depends how you use it, just last week I’ve hit 40 nanoseconds unpacking a 8 megabyte msgpack array and accessing one of its values in a hash.
As long as you only use ruby as glue code for c(++) extensions it’s pretty fast.
AFAIK with the latest JIT in some contexts pure Ruby can be faster than using C libraries, just because the VM can be better optimized and there is no overhead in moving data between the two realms.
I don't recall exactly where I read it, but I think was a while ago when they announced one of the newest JIT engines.
I recall something similar statement and I think it was from the YJIT team, they suggested that more and more people write pure Ruby rather than using C extensions because the YJIT compiler cannot see C code, it's like a black box to it, so it cannot optimize it further. Which means that in practical examples, YJIT has been able to optimize pure Ruby code to the extent that it in some cases not only beat the C extension but also surpassed it
More Ruby code means more room for optimizations that the team can identify and add to the compiler
if you want c extensions to get (de)optimized at runtime based on their usage patterns there is always ruby on graalvm from oracle.
You're thinking of TruffleRuby right? Yeah I have it bookmarked actually, it's performance is quite impressive
> As long as you only use ruby as glue code for c(++) extensions it’s pretty fast.
Another way of saying that is "as long as you don't use it it won't slow you down".
> there's no point in pretending Ruby, compared to, say, Rust
Just the thought of comparing Rubys execution speed with Rust is pointless. They are completely different languages and has their own use cases, if you care about performance, Ruby is not it. I don't think the author intended to oppose that either
I had a similar conclusion about C++. C++ takes forever to compile, but C++ is truly insanely fast, its just the compilation process is insanely inefficient. Look at Go or even D (iirc with parallel compilation). It's a night and day difference. C++ is not slow, but its compilers sure as heck are.
Edit: Another honorable mention, look at Delphi in its prime, millions of lines of code, compiles in under 5 minutes.
Neat. It automates the same kind of optimization I’ve done with ERB before (pre-compiled templates in constants). It means I don’t have to teach people were and how to optimize ERB.
To get faster this, you need to get out your profiling tools and your notebook.
Builder::XmlMarkup and strings are significantly more flexible and have many optimization points. (E.g. not escaping numbers.) They have the potential to be an order of magnitude faster. But you’ll be using that notebook to diagram where everything goes.
yo dawg meme? I think that meme's old enough to drink now
I like it, as memes go. My personal favorite:
Yo dawg, I heard yo and yo dawg like yo-yos so I put yo dawg in yo yo-yo so yo can yo-yo yo dawg while yo dawg yo-yos, dawg.