The problem is that the people getting good results with AI-assisted formal methods are the same people who get good results with formal methods without AI assistance. They then extrapolate the benefits they are getting from AI today to what it may do for others in the future, and this is where we get into trouble.
There's a lot of art to using formal methods around how to specify the system at the right level of abstraction (to make verification tractable) and how to specify the correctness properties so they can be easily evaluated. Even with AI assistance as it currently exists, users need to know formal methods well enough to at least understand the specification of the system and the correctness properties, which requires ~90% of the effort of learning formal methods in the world before AI.
But the real hope is that one day AI will be able to use formal methods correctly on its own, benefitting those who don't know formal methods. AI can sometimes do that today, but sometimes isn't good enough for people who don't know formal methods. It is certainly possible that soon enough AI will be able to do this more reliably, but then we get into the hard problem of speculating the "AI future". It is very hard to predict what an AI that can take over the art of using formal methods cannot do. Predicting that AI will be able to do that yet not be able to collect requirements and build software autonomously, or even come up with the idea for what software to build in the first place, or even replace the software's users seems arbitrary to me. In other words, if people think AI will take care of the verification letting us focus on requirement validation, my question would be, why wouldn't an AI that knows how to verify also know how to validate the requirements? For that matter, why wouldn't it also know how to replace the users altogether?
1. The study of complexity classes isn't intended to dissuade people from writing certain programs. It's intended to understand the nature and theoretical limits of computation. As far as practice goes, it can be used to show where heuristics are needed. Saying it's overrated is like saying calculus is overrated because most people don't need to use it every day. And BTW, many important problems are in classes believed to be way harder than NP (i.e. NP-complete is the easiest of the hard famous complexity classes). E.g., I've seen some people brag about some configuration language being easy to mechanically analyse because it's not Turing-complete, while in fact it's at least PSPACE-hard to analyse.
2. When there's some large set of instances of some NP-hard problem that are tractably solvable in practice (like SAT), the importance of that is that there's some non-NP-hard subset here. Indeed, SAT is FPT (fixed parameter tractable [1]), an "easier" type of NP, for which decomposition can help. In contrast, graph colouring is thought to not be FPT.
> Saying it's overrated is like saying calculus is overrated because most people don't need to use it every day.
You should stop thinking by analogy.
The article was showing the difference between mathematicians and engineers. For the mathematicians that created Computation Science, the only interesting solutions are complete solutions to general questions, whereas for engineers it's perfectly acceptable to eliminate some corner cases, thereby solving a reduced and simplified version of the general problem.
> For the mathematicians that created Computation Science, the only interesting solutions are complete solutions to general questions
Except that's not really true, which is the whole point of the finer computational classes. If many instances are far from the worst case, that tells you something interesting about the class, which is why we have things like parameterised complexity. People who think that the theory is only interested in the general case of the broad classes you learn as an undergrad are just not sufficiently familiar with the theory.
I did study a bit of complexity theory back in the day, and it does seem a bit theoretical, but not so long ago, I ran into an very prominent manifestation. To help someone, I had created a simple web page which also had an input field for the css that was applied directly to the page. To avoid flickering and weird effects because of malformed css, I wrote a regexp to check the CSS code. It's a very simple language, after all, and it didn't have to be perfect. It worked well, until at one point, a simple typo locked up Chrome for 1 minute 10 seconds. Parse time just jumped from less than 100ms to 70000ms by adding a single character. That's when you feel what exponential means.
I just want to point out that complexity theory isn't really about analysing the complexity of a particular algorithm (that's the subject of algorithm and data structure analysis, where students learn big-O notation), but about understanding the complexity of a problem independently of any algorithm (i.e. what problems couldn't be tractably solved by any algorithm).
> The article was showing the difference between mathematicians and engineers.
No. Many engineers AND mathematicians worked for a long time to get us to a stage where Amazon can solve a billion SMT problems a day. To contribute, all of them had to understand the theory this article calls overrated.
Some mathematicians certainly did, but there's a very large undercurrent in CS, as well as Mathematics more in general, of utter disinterest for applications as well as the idea that the more general a solution, the more "worthy" it is. That was really obvious from the words of the professor cited in the article.
I would to say I support that view. What is the purpose of modern science if not to discover truths you can apply universally? If you state: ‘this particular apple falls to the ground’ thats not a scientific discovery, there must be some general applicability.
2. That the worst-case is very hard usually tells you that many instances will be hard (unless you discover an easy subclass), as is the case here. And when many "natural" instances are easy, that means that the problem is more interesting than perhaps previously thought, and it requires and receives more research, not less. If most instances are near the worst case, it means you know all there is to know about the problem; when they're not, it means there's more to study.
A program that enumerates all theorems in ZFC and stops when it proves a contradiction (e.g. true = false). Encoding a program that is equivalent to that directly as a Turing Machine in merely 748 states: https://www.scottaaronson.com/papers/bb.pdf (meaning that we cannot prove an upper bound on the 748th Busy-Beaver number, but there are probably even smaller undecidable TMs).
But my favourite example (shown here in Java) demonstrates the difficulty of analysing simple, realistic programs without necessarily being undecidable:
long foo(long x) {
if (x <= 2 || (x & 1) != 0)
return 0;
for (var i = x; i > 0; i--)
if (bar(i) && bar(x - i))
return i;
throw new Error();
}
boolean bar(long x) {
for (var i = x - 1; i >= 2; i--)
for (var s = x; s >= 0; s -= i)
if (s == 0)
return false;
return true;
}
Even in this case where even the input space is finite (and so everything here is definitely decidable), we simply don't yet know whether there is some x for which foo(x) throws, let alone if we made the input unbounded by using BigInteger instead of long.
You have clearly not encountered theoretical computer scientists. They love to create all kinds of complexity classes and theorems to capture things like heuristics and approximation algorithms and other things which work in practice but not theory.
In fact that’s a big research thrust right now, to understand why many real-world SAT instances are solvable quickly while others are not, and where the threshold between them lies
I studied mathematics, and can attest that the attitude of the professor mentioned in the article is very representative of an older generation of mathematicians. Maybe the younger ones are different.
I also studied mathematics, and I can tell you that this was not my experience.
Rather, if a problem is NP-complete/NP-hard it means that we cannot expect a general fast algorithms for exactly this problem (in other words: more mathematics is required, which mathematicians of course love).
But it is absolutely known that there exist other strategies:
- Develop algorithms that work well in practice and make understanding why they work so well in practice your career.
- Find out whether there exists something that makes the instances that occur in practice different from those instances that were used in the proof that the problem is NP-complete/-hard.
- For optimization problems: develop some fast algorithm which guarantees some approximation factor.
It't not a matter of generations. There are plenty of old mathematicians who are very interested in applications and plenty of young mathematicians who are only interested in generalized abstract nonsense. It's more of a difference in personality that will always be there.
> E.g., I've seen some people brag about some configuration language being easy to mechanically analyse because it's not Turing-complete, while in fact it's at least PSPACE-hard to analyse.
Isn't even just the question of minimising the length of a regular expression PSPACE-hard or so?
> E.g., I've seen some people brag about some configuration language being easy to mechanically analyse because it's not Turing-complete, while in fact it's at least PSPACE-hard to analyse.
I don't get your point here. What analysis are you talking about?
I believe the claim usually made about non-turing-complete languages is that it is possible to prove specific properties with little to no calculations, that would be otherwise hard to calculate. For instance, the time needed to determine that an Idris program will eventually stop is litteraly 0 seconds.
> I don't get your point here. What analysis are you talking about?
Determining any kind of non-trivial property (i.e. a property that isn't true for all or none of the programs in the language).
> I believe the claim usually made about non-turing-complete languages is that it is possible to prove specific properties with little to no calculations, that would be otherwise hard to calculate. For instance, the time needed to determine that an Idris program will eventually stop is litteraly 0 seconds.
It's not the non-Turing-completeness that makes that practical. Let's take your example of Idris:
1. If a program's termination is hard to determine, then it will be hard to write it in Idris. I.e., the effort isn't gone, it's just shifted elsewhere. And if the program is easy to write in Idris, then its termination is also easy to prove in other languages (Idris effectively requires you to write a proof of termination, but you can write the proof for any language).
2. The importance of this is not as high as you think. For example, we can trivially rewrite all of the world's software in an always-terminating language (so not-Turing-complete), by changing the semantics of all programs to terminate after 2^100 steps. This will not affect the behaviour of any software, and you can see why it also won't make determining any of their properties of interest any easier.
So yes, Idris makes termination a trivial property for Idris programs, but it doesn't make the effort of determining whether an algorithm terminates or not easier (you just have to do it _while_ you're writing the program instead of after), and it doesn't, by itself, make any other property (which remains non-trivial) easier, such as "does the program terminate in fewer than 2^100 steps?"
> That is the benefit of non-turing complete languages, though.
That's simply untrue. You can write arbitrary proofs about programs in Turing-complete languages, too. In fact, most formal proofs are of programs written in Turing-complete languages.
> but languages like Rust or Idris make it simple because they force correctness early in the coding process.
Rust doesn't actually let you do that, though. In terms of the expressive power of proof, it is far closer to C than to Idris (in fact, from Idris's vantage point, Rust is almost indistinguishable from C).
> 2. When there's some large set of instances of some NP-hard problem that are tractably solvable in practice (like SAT), the importance of that is that there's some non-NP-hard subset here. Indeed, SAT is FPT (fixed parameter tractable [1]), an "easier" type of NP, for which decomposition can help. In contrast, graph colouring is thought to not be FPT.
Sorry but I need to clarify here. "SAT is FPT" does not mean anything. FPT only makes sense when you tell what is the *parameter*. Every problem is FPT when parametrized by the input size so graph colouring and SAT are FPT wrt to the size of the input (the graph and the formula respectively). What you meant: graph colouring parametrized by the number of colours is unlikely to be FPT (since it is W[1]-hard). SAT is FPT for many parameters such as treewidth (of the formula). Oh, and btw, graph colouring is also FPT when parametrized by treewidth (of the graph).
I take your correction re graph colouring, but the rest is pointlessly pedantic (especially "FPT in the length", which is trivial and so typically excluded from the definition).
Now, I don't know if SAT being FPT (in a parameter of interest) has anything to do with the surprising ease of many "natural" instances (even with many variables), but my point was that when many instances are easy, that doesn't mean that the class is irrelevant, it just means it's more interesting (as there's obviously a tractable subclass, albeit one we haven't yet defined succinctly).
Late answer but I need to be pointlessly pedantic again, because it is apparently what it is called to correct something that does not make sense. That said, I agree with your way of reformulating your point.
To first address the pedanticness: "FPT in the length of the input" is not excluded from the definition of FPT even if it is not an interesting case. Excluding these cases from the definition would make the theory uselessly complicated.
Now, back to the point of my first comment. It was mostly to show that you cannot use "FPTness" as an argument to justify that SAT is "easier" than graph colouring. I used "length of the input" to give an easy counter example, to show that FPT is not "an easier type of NP". It never has been. It is a way of understanding and measuring the complexity of the problem finely, to isolate hard parts of the input from the rest. What the W[1]-hardness of graph colouring parametrised by the number of colours tells you is that this parameter is not a relevant parameter and that's it. Graph colouring is FPT for many relevant graph parameters such as treewidth, clique-width (SAT is not even FPT wrt clique-width, but again, you cannot really use the concept of FPT to compare problems).
You will have a hard time using theory to justify that SAT is easier than any other NP-complete problem, because from the theory point of view, it is the hardest NP-problem you can get. Unless something unexpected happens in complexity theory, SAT cannot be efficiently solved in randomized time, SAT cannot be solved in sub-exponential time, SAT is complete under parsimonious reductions so you can basically take any NP-complete problem and build a CNF formula whose models are isomorphic to the solutions of the original instance etc.
The success of SAT solvers does not mean that SAT is easier than other NP-hard problem. It means that many combinatorial problems we need to solve in practice are "simple enough" that using the incredibly optimized smart way of bruteforcing the solution with a CDCL SAT solver is good enough. Now, take a cryptographic instance, translate it into a CNF formula and call a SAT solver, I doubt it will shine. Despite many attempts, this behaviour has never clearly be explained by the fact that the complexity of CDCL is FPT wrt a parameter that is small in industrial instances.
> You will have a hard time using theory to justify that SAT is easier than any other NP-complete problem
"Easier" perhaps isn't the right word (in terms of reducibility, it certainly isn't), but FPT is just a way to demonstrate that NP-complete problems can be different from each other in ways that matter (with respect to "natural" instances).
> Now, take a cryptographic instance, translate it into a CNF formula and call a SAT solver, I doubt it will shine.
Exactly. And researchers are very much interested in these differences that arise in natural instances and are hidden by "crude" reduction.
> It's intended to understand the nature and theoretical limits of computation.
Not in a general sense, at least for standard complexity theory. It only deals with a very specific model of computation. Anyone with a sufficiently solid grasp of metamathematics intuitively understands that the distinction between solve and verify is nothing but a description of how badly matched our foundations are for the structure we're trying to view.
... This is the second time today I've posted about foundations like this.
Standard complexity theory focuses on answering questions when our substrate behaves like a Turing machine with multiple tapes.
Consider it like this, if the answer is in our system's axioms, we don't have to do anything. In a trivial sense that means we're just given the answer table, but it's also true if our substrate matches the model of computation its simulating. IE for an SLD-Resolution machine, running an SLD-Resolution object language, unification is worst case O(1). This is a degenerate case of course, but it's an example of something that's not realizable on a Turing machine's semantics where the worst case is in... EXPTIME? It's not great.
The more we treat our substrate like building blocks, and less like a holistic oracle, that changes our complexity landscape. Complexity theory was never about studying that whole landscape.
You might want to say CT is pragmatic and focused on realizable machines. There are two problems with that:
1. There's nothing special with the baseline used for complexity theory other than its familiarity. Reality is our ultimate substrate. The universe is not Turing tape. There is absolutely no serious basis upon which an argument against substrates can be made, especially with how little we know and understand about the universe.
2. Complexity theory isn't so pragmatic to only study the finitely bounded, which also changes everything. There seems a very tight upper bound on information in the universe. Even studying up to it as a limit is decidedly not pragmatic in the slightest. This is perfectly fine of course, the problem only enters in when we want to be "pragmatic" on some things, but not others.
I also want to clarify: There are higher orders of complexity theory that have generalized a lot of its concepts, even into hypercomputation which is cool, but then there's another problem I didn't mention. Complexity theory still isn't about what he said. It quantifies that distance between prove and verify, but it doesn't study the set of all those distances and how they arise. It just quantifies them one at a time and has only a limited number of things to say beyond that. What he described is simply mathematical logic.
We actually understand quite a bit about the universe and the kinds of computers we can build. People also think about computers in speculative physics scenarios, eg closed timelike curves can be used to solve pspace complete problems.
I don’t think you can plausibly argue that complexity theory‘s base assumptions are a bad choice, at least not in the sense that you would assume that you can build exponentially more powerful computers in the physical universe. In fact, concerns about energy densities, limited amounts of matter, and the speed of light make it more difficult than typical machine models assume.
> Standard complexity theory focuses on answering questions when our substrate behaves like a Turing machine with multiple tapes.
This is not true. Complexity theory very much looks at complexity under different models (alphabet size, oracles, circuits). It's just that often (e.g. in the case of alphabets), there is a reduction of known complexity between two models.
> It quantifies that distance between prove and verify, but it doesn't study the set of all those distances and how they arise.
> there is a reduction of known complexity between two models.
Every single time I've seen, for example, the lambda calculus be assigned cost semantics, it usually looks like what you would expect out of a Turing machine's simulation of it. Often times, they're explicit about it: https://www.sciencedirect.com/science/article/pii/S030439750...
For me, I can't accept that this is the criteria of "reasonable." Especially not for abstract theory.
I did try to indicate I'm mostly talking about standard complexity theory, the stuff you'd encounter on the surface level of the field. I'm not an expert in CT, but I do know enough to know what Landauer's principle is (and that it's been plausibly challenged.) I also know there's some crazy stuff in there, like descriptive complexity theory's link between Existential SOL and NP-Complexity.
Do you have any complexity theory papers that deal with this specifically? I've only ever seen that kind of work done in mathematical logic. Genuine interest in reading the CT approach.
> it usually looks like what you would expect out of a Turing machine's simulation of it
You could assign it any cost model you want. Often this doesn't make a difference (as speedup theorems and other "distracting details" mean that most classes are intended to be separated by exponentials), but it is true that people are typically more interested in cost models that are more relevant to the physical universe (although number of reductions is very much the cost of focus in proof complexity). Indeed, relevant discoveries in physics yield corresponding computational complexity research, as in the case of quantum complexity (https://en.wikipedia.org/wiki/Quantum_complexity_theory).
> Do you have any complexity theory papers that deal with this specifically?
A Google Scholar search for "proof complexity" will show you many papers as well as a number of books.
The point that what gets published adapts to the state of the art is fair. Personally I think its adaptations for ANNs and interaction nets are more interesting than the quantum ones. Once it has to account for the topology of the computer, or deal with non-atomic and continuous substrates, it changes a little bit. But this is orthogonal. The point wasn't that you can't make it adapt, I hope it doesn't come across like I'm dismissing the field and its importance. What I'm trying to say rather specifically is that it's not it's not about studying the nature of computation. It lives at a different level of theory than that. It might just be a disagreement on the article being used. I think it would be fair to say it studies a nature, just not the nature, but that would be true of all fields of CS.
I would say that it very much is about the nature of computation, but computation itself has always been tied to the physical. The physicality (even hypothetical physicality) is exactly what separates theoretical computer science from pure mathematics. Even the two primary resources that complexity measures - time and space - are tied to physical quantities.
The distinction is certainly a fair one to make. The etymological root of computation is "done with mental labor", by way of "to clean a financial ledger" (although weirdly, this isn't the real root. Putare is botancial pruning. Computare is a metaphor created after it spent some time on the semantic treadmill.) Grounding it to the physical is perfectly sane, because that's historically how it's been used.
That being said, a counterargument to press against this is that complexity theory doesn't restrict itself to physical or hypothetical physicality in its totality. As I mentioned, there are swaths of complexity theory work which bound quite far afield. The higher orders of the field are decidedly not-physical at all (and pedantically, hypercomputation isn't strictly computation). Of course even beyond this, we're still not studying the nature of computation, we're studying the cost. While you might say that of course these things are tied thanks to physicality (I wouldn't agree that they're equivalent on this basis, but I don't think that's an interesting semantic argument), I did also mention Landauer's principle being plausibly challenged, which is further problematic for conflating the two. Computation being reversible where entropy isn't doesn't explode complexity theory, but it does drive a wedge between information in a computer-sense and information in a thermodynamics sense. At that point, we don't have the claim in the first place, it's just a false friend. Something to consider.
I'm not saying computational complexity restricts itself to the physically realisable, I said it's tied to it and so there will be more papers on models that have some correspondence with physical reality.
> we're still not studying the nature of computation, we're studying the cost
That is one way of studying the nature of things, sort of like the use of the Hamiltonian in physics, especially if you're interested in problems and classes of problems, and their broad similarities and differences via reductions, rather than in specific computations.
> The higher orders of the field are decidedly not-physical at all (and pedantically, hypercomputation isn't strictly computation).
Hypercomputation isn't really a big part of the main thrust of complexity theory, but (computable) oracles do very much play an important role in complexity theory, even in its lowest complexity classes, such as in the relativisation barrier, which shows that some proof techniques cannot separate P from NP.
> I did also mention Landauer's principle being plausibly challenged,
That's not complexity theory, at least not the standard theory, which treats time and space (or circuit size) more abstractly than concrete physics. There are, however, theoretical reversible models, just note that they don't yield different "classic" deterministic complexity classes (i.e. they do not yield exponential differences).
Again, all fair. To simplify, my points are to demonstrate a confounding nuance, simple counterexamples where the arguments don't hold.
> Hypercomputation isn't really a big part of the main thrust of complexity theory
I don't mean to imply that it was, though the results are actually relatively important elsewhere.
> That's not complexity theory, at least not the standard theory, which treats time and space (or circuit size) more abstractly than concrete physics.
More, but not totally abstractly. Steps and cells being vacuous primitives, they're not literally space-time, but within orthodoxy there's absolutely a partial morphism that's implied. That's why they're named like that. You are supposed to have them live close together in your head.
>There are, however, theoretical reversible models, just note that they don't yield different "classic" complexity classes
I know that some don't, but for example quantum models to use your own example, while not technically reversible in the absolute sense, do possess some reversibility capability and do derive different complexity classes.
I think it's very obvious that there should be reversible computational models which yield different complexity classes from the typical ones. To me for a field to qualify as a study on the nature of computation, it should probably be able to design one totally a posteriori, if in a higher order language. Complexity theory might be invoked in such a construction, but it's not the one doing the building. It's one of many in an orchestra.
Here's a question I have, since you do seem pretty well versed on CT. Universal quantification over complexity classes of first-order systems, used anywhere in the abstract?
> Universal quantification over complexity classes of first-order systems, used anywhere in the abstract?
I'm not sure what you mean by "complexity classes of first-order systems" and by "in the abstract".
But it seems like you're asking about the intersection of computational complexity and formal systems, and there's definitely work there. I already mentioned proof complexity, which analyses the number of deduction steps needed to prove something in various formalisms, and there are famous undergrad-level examples, such as TQBF (https://en.wikipedia.org/wiki/True_quantified_Boolean_formul...). But an intersection that is of more interest to me, as I'm interested in software correctness, is that of the model-checking problem.
Now, many people are confused whenever the model checking problem is discussed, because they confuse it with model checkers, which are a set of algorithms intended to solve the problem, but complexity theory is typically interested in the inherent difficulty of answering problems regardless of the algorithm used to do it. So the model checking problem is that of determining whether a formula in some formalism implies another formula, and its inherent complexity exists regardless of whether this question is answered via a formal proof or by some technique involving the logic's semantics. In the context of software verification, the model checking problem is that of determining - by whatever means - whether a program satisfies some non-trivial property.
Philippe Schnoebelen has some papers on the model checking problem in temporal logic (https://lsv.ens-paris-saclay.fr/Publis/PAPERS/PDF/Sch-aiml02..., https://lsv.ens-paris-saclay.fr/Publis/PAPERS/PDF/DLS-jcss-p...). One of his findings that I've found most interesting with regards to programming is that programming languages cannot, in general, make answering the question of whether a program satisfies some property any easier. This result is surprising. The reason is that without a programming language, we could describe a program as a huge state transition graph (this is called a Kripke structure). In that representation, it's been proven that verification is linear in the number of states, i.e. there is no general approach that is faster than brute-force. Now, the size relationship between a program written in a programming language and its Kripke structure is easily exponential or more, so if there were no algorithm that's better, in the worst case, than a brute force of the Kripke structure, then obviously verification is intractable in the size of the program. However, the number of Kripke structures of size N that have a succinct representation in some programming language is far smaller than the total number of Kripke structures of size N. So it could have been the case that analysing programs would have been easier than analysing their Kripke structure (while ignoring their representation in the language). But Schnoebelen proved that this is not the case.
He also proves that program decomposition (and verification of each component separately) cannot, in general, make verification any easier (i.e. the model checking problem isn't FPT in the number of program components).
These results are far more recent than the hopes expressed in the seventies and eighties that we'll be able to prove the correctness of all/most/many programs we write, and indeed, even though the results talk about the worst case, what we've seen in the last few decades is that the power of program verification indeed behaves more like the worst case than something far from it. The gap between the size of programs we can verify and the average size of programs we write has only widened (what saved the day has been the effectiveness of unsound methods, but that's a whole other discussion).
> you just do all that manual work with environment variables
You really don't anymore. For the past several years, Java's GCs mostly pick the right settings automatically, except for heap size, which will be taken care of soon (https://openjdk.org/jeps/8377305). The reason heap size isn't automatic is that with moving collectors it determines the CPU/RAM tradeoff, and doing that in a more natural way isn't trivial, but we have the algorithm now and will merge it soon.
> or making sure to "pick the right collector for the job"
There are really only four options, most of which are easy to choose among: Parallel for batch jobs where only throughput matters, ZGC for interactive applications where latency matters a lot, and then consider either G1 or Serial if there's a problem with those choices.
As someone who's worked for a long, long time solving manual memory management issues, the amount of effort required isn't just in a different ballpark, but in a different city. Sure, spending a few hours a year to reconsider your settings isn't nothing, but it isn't even remotely in the same category of pain with manual memory management (or even automatic memory management, but with malloc/free underneath).
Concretely, what are current tail latencies, worst case?
Ten years ago, “rewrite in C++” was definitely easier than getting the Java GC to stay up under server load.
Most servers I work with run on big machines and are the only process, so figure a 100-250GB heap that lives for months, all async, small requests, so insane amounts of Future and String allocation spam.
Optimizing that stuff away in Java is harder than writing Rust, so assume idiomatic Java.
Also, is there any work on statically enforcing data race freedom in Java? That’s a bigger rust selling point than memory safety for me. I think swift has done some interesting work in that space. It would be nice to get those sorts of safety properties without manually writing borrow checker annotations.
As the OC, I think my view is somewhere in the middle - I am neither as optimistic about it being "great now" nor do I think that "rewrite in C++" 10 years ago was easier.
My reason for disagreeing with the former view is that improvements in physical RAM available and tendency towards smaller workloads have allowed many Java (or other GC runtimes) to essentially "fix their problems because hardware got better". So you can waste more RAM, waste more cycles, but "it doesn't matter", and likely it is fine in many cases - but it's is not the same thing as claiming the GC algorithms are responsivle for that outcome. We have been 3 years away from GC solving memory management for at least 30 years.
My reason for disagreeing with the latter view is that for those who don't have 100-250 GB long-lived heaps (or whatever the contemporary version of that is), the pain level is far lower than rewriting in C++ or Rust, or likely the pain level of hiring enough engineers who can do either. It's a completely different engineering culture.
Just to be clear, the main reason for the use of moving collectors in the first place is to waste less cycles on memory management (otherwise we wouldn't use them). They exist to serve as an optimisation.
> We have been 3 years away from GC solving memory management for at least 30 years.
It's now 3 years in the past (since Generational ZGC); e.g. see https://netflixtechblog.com/bending-pause-times-to-your-will.... Of course, it doesn't solve all imaginable memory management issues, but in practice it makes it a non-issue for a large class of interesting and very common programs.
Maybe I should've mentioned at the start that I've implemented several GCs and worked on several Java VM implementations, so I am generally familiar with the tradeoffs between GC algorithms and other runtime details.
Even in the very positive blog you linked, you see statements like
* "ZGC has a fixed overhead 3% of the heap size, requiring more native memory than G1. .." and
* "Reference processing is also only performed in major collections with ZGC. We paid particular attention to deallocation of direct byte buffers, but we haven’t seen any impact thus far. This difference in reference processing did cause a performance problem with JSON thread dump support, but that’s a unusual situation caused by a framework accidentally creating an unused ExecutorService instance for every request."
This was my point about how this sort of thing is a type of manual memory management.
As for waste more RAM, waste more cycles wasn't a statemnt about whether a particular GC is better-performing for certain situations, but that the overall improvement likely has more to do with improvements in CPU speeds and RAM size, than the latest GC version (which tends to simply make a different set of engineering tradeoffs).
You're right that managing any resource that isn't just Java heap memory requires manual management, but it isn't what we normally mean by memory management. The two have been somewhat tied together traditionally through reference processing (in the sense of reference queues), which is generally something we now discourage in Java programs, and may be deprecated and removed altogether at some point. It's traditionally been used as a convenience. The framework used in that post is, indeed, based on an old library that relies on manual or reference-processing-assisted management of non-heap resources.
> but that the overall improvement likely has more to do with improvements in CPU speeds and RAM size, than the latest GC version (which tends to simply make a different set of engineering tradeoffs).
Well, the biggest improvement has been the creation of a new "pauseless" collector, ZGC, with a novel GC algorithm (at least for OpenJDK), which does _zero_ GC work in STW pauses, i.e. no scanning, no marking, no moving. In particular, even roots, including stacks, are processed entirely concurrently with the program. The main practical impact of that has been saying goodbye to GC pauses, and getting low latency, that is perhaps even more predictable than malloc/free (and obviously, still has higher throughputs in a large class of interesting programs). The tradeoff is the usual footprint tradeoff, which is the core of moving algorithms, as well as more CPU cycles compared to STW collectors (but again, still less than malloc/free in many programs). The additional CPU can, of course, be compensated for with an even larger heap.
The general idea is to use RAM chips as hardware program accelerators, but in the past latency was also something you had to sacrifice, and this is no longer the case today.
It definitely was easier for the projects I worked on, but they are exactly the use case where the heap is long lived and most of the machine.
I’ve also worked on systems with lots of small processes, and the operational issues that creates dwarfs GC problems: It takes one middle tier machine, and adds 64-128 network boundaries, and also creates an extremely difficult static memory allocation problem.
I know people do it anyway, but it’s rare that they can articulate a decent technical reason for it, and it wastes something like 90% of the hardware (even in carefully optimized code bases / deployments).
> Concretely, what are current tail latencies, worst case?
Well under 1ms for ZGC (to the point that OS-caused hiccups are of similar magnitudes).
> Ten years ago, “rewrite in C++” was definitely easier than getting the Java GC to stay up under server load.
Both could have been hard in some cases, but open-source "pauseless" GCs are only 3 years old (and all of the JDK's GCs are nothing like what they were ten years ago).
> Optimizing that stuff away in Java is harder than writing Rust, so assume idiomatic Java.
Quite the opposite. Performance issues due to memory management are, in practice, more serious in Rust than they are in modern Java.
> Also, is there any work on statically enforcing data race freedom in Java?
There isn't much demand for that atm. If we see growing demand, we could prioritise it.
In rust, I usually just make sure stuff is not Box<>, and try to reuse buffers. That generally gets the memory allocator completely out of the way (except for async).
The remaining allocator performance problems are mostly due to it zeroing allocated memory unless I use unsafe. Is java able to stackify most new Object calls and elide default initialization of object members these days?
I’m surprised to hear there is no demand for compiler enforced/facilitated thread safety in Java. That was a major pain point in all the Java code bases I’ve worked with in the past, and is a headline safety feature for rust (which goes even further and enforces aliasing rules) and JS. Could you be seeing selection bias in your user base?
> Is java able to stackify most new Object calls and elide default initialization of object members these days?
Escape analysis in OpenJDK will stack allocate values where it can show it is safe to do so. Project Valhalla is also reducing the memory footprint of objects.
As for thread safety, that is more of a language concern than a runtime one. Amongst JVM languages Scala is leading here AFAIK. Its "capture checking"[1] provides thread safety (e.g. [2]) and actually covers escape analysis as well. On Scala Native (the native code backend for Scala) capture checking can be used for safe stack allocation and safe arena allocation.
> In rust, I usually just make sure stuff is not Box<>, and try to reuse buffers. That generally gets the memory allocator completely out of the way (except for async).
You say "just", but this is easy when programs are small. The problem is that this gets harder and harder and harder as programs grow large (the whole point of the JVM's design was to address the performance issues that plague large C++ programs). E.g. someone who works at one of the world's largest tech companies just told me that they have problems with Rust programs spending 30% of their CPU on memory management even when they're as small as a couple hundreds of thousands of LOC.
> Is java able to stackify most new Object calls and elide default initialization of object members these days?
No, the general idea is to just make memory management efficient (although some objects are "stackified" and the compiler will elide zeroing when non-defaults are passed to a constructor). Now, I say "just", but this used to come at the cost of GC pauses and larger footprint. Now it only comes at the cost of a larger footprint.
But there is a definite choice here when it comes to performance. Low level languages give you control that means performance is attained through manual effort. Java takes away control to improve effort-per-performance. Roughly speaking, these tradeoffs mean that when programs are small and the extra effort is manageable, low-level languages are hard to beat, but when programs are large, it is Java that is hard to beat.
> I’m surprised to hear there is no demand for compiler enforced/facilitated thread safety in Java. That was a major pain point in all the Java code bases I’ve worked with in the past, and is a headline safety feature for rust (which goes even further and enforces aliasing rules) and JS.
This used to be a bigger problem when locks were the main mechanism for sharing data among threads. Now, with the wide selection of concurrent data structures, such problems don't occur as much. I'm not saying they don't occur at all, just not frequently enough to become a major priority.
Also, safe Rust's data-race freedom comes at the cost of requiring unsafe for benign races, which are not uncommon in concurrent algorithms (i.e. it excludes even "good" races). This may be fine in languages whose view on performance is "with enough effort you can get good performance", but, as I said, Java is about making more "naive" programs fast with little effort.
It is fair that there are many ways to be slow in any number of programming languages. I'm surprised to hear "Rust programs spending 30% of their CPU on memory management even when they're as small as a couple hundreds of thousands of LOC", although I can visualize some unique workloads where that's unavoidable irrespective of language & runtime.
You say unavoidable, but moving collectors are designed to reduce CPU at high allocation rates by increasing the heap size. Generational moving collectors have a pathological case - a high allocation rate of long-lived objects - but that's quite hard to get yourself into by accident. Their main downside (besides the inherent increased footprint) used to be unpredictable long pauses, which could have a very high impact on tail latencies, and that's just gone with ZGC.
Generational moving collectors are a very powerful memory management optimisation, but because they necessarily require some "interesting" FFI layer between the ordinary heap and any passing of pointers between the program and the hardware/OS - the very thing low-level languages are designed not to have - this is a powerful, general optimisation (not perfect, but extremely useful in a wide class of programs) that is not available to low-level programming languages. And it's not the only one, BTW. JIT compilers are also designed for "global" average-case optimisations at the cost of precise low-level control over the worst case, which is another thing that low-level languages trade away.
In the most simplistic way, I would say that the precise control that low-level languages are all about helps their performance when programs are small (and can be manually optimised globally) and hurts their performance as programs get large. They have to give up on some optimisations that come at the cost of ceding low-level control, and that includes moving collectors.
When I say "unavoidable" I mean exactly things like your "a high allocation rate of long-lived objects (as one example), where the work that has to be done requires a significant number of CPU and memory cycles spent on memory management (manual or automatic).
When I say tradeoffs", I mean exactly things like "inherent increased footprint", or my earlier "wasting RAM" point.
At this point, I'm not sure that we're disagreeing about these details, but rather what we make of them... I view them as just years of continued tuning of tradeoffs (heap size vs CPU cycles vs memory cycles), while you them as major breakthrough that makes garbage collection much more desirable. Is that fair?
A high allocation rate of long-lived objects is not easy to do. Object "death" rate has to equal the allocation rate, so a high allocation rate of long-lived objects means that somehow you get to allocate, say, 1 GB/s of data that is kept for a long time and is discarded at a rate of 1 GB/s.
> When I say tradeoffs", I mean exactly things like "inherent increased footprint", or my earlier "wasting RAM" point.
Well, that is a real tradeoff, but there's a reason why it's a very attractive one for a huge class of applications. There are two ways of looking at this, which amount to the same thing:
1. Because both RAM and CPU are needed for computation, what matters isn't each of their utilisation values separately but only the more constraining or impactful of the two.
2. Because CPU is needed to use RAM, every CPU cycle you spend effectively takes away some other program's ability to use RAM.
This means that for any amount of CPU utilisation, there is some amount of RAM that is effectively free (i.e. has no additional impact), and the more CPU a program consumes, the more RAM it can consume without it making additional impact. It's easiest to see in the extreme case of a program using 100% CPU: no other program can make progress, and so it doesn't matter how much of the available RAM your program is using - it effectively captures all of it whether it uses it or not. But this scales to any amount of CPU utilisation (not quite linearly). What wastes RAM is not using that "free RAM" to reduce the dominant resource, CPU. And what further determines the RAM/CPU "exchange rate" is the RAM/CPU ratio offered by the hardware, which is more RAM-heavy than some appreciate (it is very hard to find a metal or virtual deployment with less than 1 GB of RAM per core these days - taking into account partial cores in virtual machines - outside of embedded devices).
This means that if you have a memory management algorithm that uses more RAM to help reduce CPU as CPU utilisation rises, that's usually a good thing. And moving collectors work exactly like that. The heap overhead in a generational moving collector is only a function of the allocation rate, and a high allocation rate also means high CPU usage.
My colleague, the main developer of ZGC these days, gave a keynote about this very subject at ISMM: https://youtu.be/mLNFVNXbw7I
> I view them as just years of continued tuning of tradeoffs (heap size vs CPU cycles vs memory cycles), while you them as major breakthrough that makes garbage collection much more desirable. Is that fair?
I say that for many years, the main practical, most "felt" tradeoff of moving collectors has been their STW pauses. With pauses eliminated, there is a qualitative change in the attractiveness of moving collectors, making them more appropriate than other memory management techniques for an even broader class of applications than before. Previously, applications that were very sensitive to tail latencies didn't want moving collectors; now, the tail latency is no longer an issue (unless your application's tail latency tolerance is such that a realtime OS is needed). In other words, I'm saying that moving collectors' most impactful tradeoff is now gone.
Does it provide hard latency bounds like Azul does (did?), and are they lower than disk/network latencies on modern hardware?
I moved to c++/rust years ago because those languages do, and tens of milliseconds matter for network services. At the time Java could pause for 10’s of seconds, which was 1000x worse than waiting for a spinning disk to seek.
These days, disks are 100s micros to single digit millis, so I guess if Java GC is finally working 30 years after they “fixed” its performance problems, then I’d want to be able to tune ZGC to not pause the app for more than ~ 500us, max.
This article is from last year, but suggests they’re still off by an order of magnitude:
If you hammer a 100GB-1TB heap in steady state with small allocations for, say, a month at 100% CPU, does it eventually do the typical Java thing, where a major compaction takes the process down for seconds or even minutes, or does it just slow down application requests so it can keep up with load?
> If you hammer a 100GB-1TB heap in steady state with small allocations for, say, a month at 100% CPU, does it eventually do the typical Java thing, where a major compaction takes the process down for seconds or even minutes, or does it just slow down application requests so it can keep up with load?
No. Every garbage collection in Java relocates objects. Compared to malloc, memory fragmentation in long-lived processes is less of a concern. Freelists track only large segments of available memory. The allocator reserves a segment per thread and simply advances a pointer. Small short-lived objects are never visited by the collector. Instead, live siblings are relocated elsewhere before the entire segment is reclaimed.
The above holds true for all of the collectors. The difference is how they deal with concurrent changes to object pointers by the application. Generally, stopping the world uses less net CPU than the memory barriers required by G1GC and ZGC, but most applications are willing to provide more memory and CPU in exchange for shorter pauses.
> Does it provide hard latency bounds like Azul does (did?), and are they lower than disk/network latencies on modern hardware?
Yes and yes (although we need to be more precise when we talk about latencies; see next paragraph).
> These days, disks are 100s micros to single digit millis, so I guess if Java GC is finally working 30 years after they “fixed” its performance problems, then I’d want to be able to tune ZGC to not pause the app for more than ~ 500us, max.
1. You don't need to tune it. The algorithm simply doesn't collect garbage in stop-the-world pauses.
2. Hiccups are sporadic. They should not be compared to the average latency of normal operation. The relevant question is, is ZGC introducing longer hiccups than those a non-realtime kernel would, and the answer is no.
> This article is from last year, but suggests they’re still off by an order of magnitude
The article doesn't measure GC pauses when it shows latencies (it says: "With ZGC on the other hand, the longest GC pause time observed is ~50 microseconds"). It measures the response latencies of some service. Note that allocation stalls also occur with malloc, it just isn't reported conveniently.
Of course, one of the greatest advantages of moving collectors still applies: Under high allocation rates, moving collectors (but not malloc/free!) allow you to compensate for increased CPU spent on memory management by increasing the heap (i.e. if your allocation rate doubles, you can increase the heap and keep the CPU cost of memory management the same). In the past, this advantage translated to higher throughputs compared to malloc/free, but suffered from GC pauses. Those pauses are gone today.
> If you hammer a 100GB-1TB heap in steady state with small allocations for, say, a month at 100% CPU, does it eventually do the typical Java thing, where a major compaction takes the process down for seconds or even minutes, or does it just slow down application requests so it can keep up with load?
No, it does not. You could, of course, construct some pathological cases where you'd have a high allocation rate for long-lived objects which would result in high CPU utilisation by the GC, but it's easier to get into pathological malloc/free cases in C++ (or Rust) than with ZGC. Let me put it another way: no matter your memory management strategy, it's possible to overwhelm it, but the likelihood that a real, "naive" program would overwhelm a malloc/free allocator is higher than it would the JDK's GCs.
> Go, Java, C#, and Python use garbage collectors. This makes them easier to use but slower and less predictable.
It does not. The term "garbage collectors" covers a whole spectrum of algorithms, some might slow you down (though not for the reason you may think) while others were invented to speed up memory management beyond that of C++, in exchange for other tradeoffs. Python's (mostly) refcounting GC is actually closer to C in its memory management overhead than to either Go or Java. It's also not what makes Python slow. Go uses a mark-and-sweep collector to find a balanace between speed, FFI, and footprint. Java uses moving collectors, which are faster - and some of which are even more predictable - than memory management in C++. That's because Java aims to offer better performance than C++ in large concurrent software, where low-level languages tend to suffer from various overheads due to their requirement for low-level control (Java trades off some performance in smaller programs, but mostly it trades of startup time and footprint). Moving collectors (but not refcoting collectors or mark-and-sweep collectors) are an optimisation over free-list approaches, not a compromise for convenience.
So it is true that slow programming languages tend to use some kind of GC, but that's not what makes them slow, nor does it make the super-fast languages that also use a GC (often of a very different kind) any slower. The range of languages that use GCs covers everything from the super slow to the super fast.
I don't understand how you can claim that using a GC does not make a language slower and less predictable.
Running a GC takes time, pollutes the cache, and is often run at an unpredictable time. Sure, the GC is not necessarily the SLOWEST thing about the language (python), but it's not helping, either.
Yes, but for a moving collector that's less time than it takes to run malloc and free. The interaction of a moving collector with most object is bump allocation when they're allocated (similar to stack allocation) and... that's it. The GC never sees them again, scans them again, or is even aware of their existence (moving collectors don't have a free operation). Overall, moving collectors (but not other kinds of GC) reduce the work of memory management compared to malloc/free.
In low-level languages we try to avoid doing a lot of malloc/free not because heap memory management is slow in general, but because that approach to memory management is slow. Moving collectors are an optimisation designed to make heap memory management fast, but it requires that (nearly) all pointers be movable, something that low-level languages can't do because they have constraints that are more important to them than speed (you can't interact with the OS or hardware directly, i.e. without an FFI API, if your pointers are movable, and such direct interaction is the point of low-level languages).
That moving collectors (NOT the GC Python has; NOT the GC Go has) can, in principle, make heap memory management cheaper than stack allocation has been well known since the eighties. But until recently they had excellent throughput (somewhat similar to arenas) but potentially long pauses. It was only recently that they were made "pauseless".
> and is often run at an unpredictable time
How much work malloc and free need to do is also unpredictable, and a modern pauseless moving collector like ZGC spreads the work needed for memory management more evenly than malloc and free.
> Sure, the GC is not necessarily the SLOWEST thing about the language (python), but it's not helping, either.
There is very little resemblance between CPython's GC and Java. Python's memory management is closer to C's than to Java's. GCs cover such a wide spectrum of algorithms that it doesn't make sense to talk about them as a single category as far as performance tradeoffs are concerned.
I've written two separate moving collectors for dynamic language runtimes, as well as done significant work in realtime 3D graphics, and what you're saying mostly smells like bullshit.
> The interaction of a moving collector with most object is bump allocation when they're allocated (similar to stack allocation) and... that's it. The GC never sees them again, scans them again, or is even aware of their existence
I mean .. sure, but, a copying GC eats somewhere on the order of 10% of your total memory bandwidth just copying shit around. I can guarantee that if you use sane allocation strategies (arenas & freelists, pools, whatever) you spend <1% of your total system resources fucking around with memory allocation.
> Moving collectors [...] requires that (nearly) all pointers be movable, something that low-level languages can't do
Completely false. You have to do some manual bookkeeping in C++, Rust, Zig, whatever, but you can do it, and in fact many commercial GCs do (V8 is a good example).
...
The rest of what you said is just empty-sounding claims that I'm not going to address. I looked at the single paper that you linked in another comment, from the 80s, which is hardly relevant on modern hardware.
Please, if you're going to make the claims you're making, back them up with hard evidence. I've looked, and the overwhelming majority of papers out there claim that GCs are slow, memory hungry and, generally, a waste of time.
> I've written two separate moving collectors for dynamic language runtimes, as well as done significant work in realtime 3D graphics
That's nice. I have ~25 years of experience with large C++ software, including hard and soft realtime systems, and I now work on the JVM.
> I mean .. sure, but, a copying GC eats somewhere on the order of 10% of your total memory bandwidth just copying shit around
Good moving collectors are designed to copy very little. That's the entire purpose of generations.
> I can guarantee that if you use sane allocation strategies (arenas & freelists, pools, whatever) you spend <1% of your total system resources fucking around with memory allocation.
In theory. When you're in charge of a >2 MLOC C++ system, maintained by a large team for well over a decade, you find that these optimisations are very costly. Plus, you commonly find large C++ software that needs sophisticated malloc/free allocators for acceptable performance (BTW, some of those allocators are almost the same size, in LOC, as that of ZGC, probably the world's most sophisticated moving collector).
That's why, where I used to work and oversee large projects, we migrated pretty much all systems (mostly soft-realtime defence software) to Java from C++ - for better performance. One of the goals of the JVM was to tackle the familiar performance issues that plague large C++ programs.
And BTW, using arenas isn't so easy when programs get large and sprawling, or even in general. Zig definitely makes that much easier, though.
> Completely false. You have to do some manual bookkeeping in C++, Rust, Zig, whatever, but you can do it, and in fact many commercial GCs do (V8 is a good example).
What I said is completely true, but you may have misunderstood it. You can, of course, combine moving collectors with things that expect stable ones in the same process. In fact, you absolutely must, because at some point in the stack you need to talk to the OS and/or hardware, and they expect stable pointers. But you have to have some distinct FFI layer between the two, and the entire point of low-level languages is to be at the lower level.
> Please, if you're going to make the claims you're making, back them up with hard evidence. I've looked, and the overwhelming majority of papers out there claim that GCs are slow, memory hungry and, generally, a waste of time.
I don't know what you've read, but that is very clearly not the consensus among memory management experts. My "claims" are pretty common industry knowledge, and why the majority of performance-critical large software has migrated away from low-level languages over the past couple of decades, and the trend continues. I'm not trying to change anything, I'm just explaining why the industry is doing what it's doing to those who may not be familiar with large, long-lived software.
It's funny, but 25 years ago, the people who doubted the amazing performance-per-effort of moving collectors and JITs were those who (like me) had not used those technologies and were mostly familiar with low-level languages. These days, it's the people who have little experience developing and evolving large and complex software in low-level languages (TBF, there's much less such software written in low-level languages these days) that believe the low-level languages are inherently fast.
Having said all that, when programs are relatively small and/or not very concurrent, the effort required to match or beat Java's performance in a low-level language through careful manual optimisation is sometimes worth it. In large software, it gets harder and harder.
> I can guarantee that if you use sane allocation strategies (arenas & freelists, pools, whatever) you spend <1% of your total system resources fucking around with memory allocation.
I just spoke with someone on the performance team at one of the world's largest tech companies who told me that some of their larger Rust programs spend 30% of their CPU on malloc/free. Of course, it's possible in principle to reduce this given enough effort, but this is identical to the experience we've had with C++ for decades: When programs are small, it's easy to get good performance, but as they get larger, the areas where low-level languages have intrinsic inefficiencies (such as dynamic memory management) tend to become more pronounced in practice and the programs are not so easy to optimise.
The quoted text says "slower". It does not claim that GC makes those languages slow overall.
Are you arguing that GC is not inherently slower than other memory management strategies (e.g. the Rust approach)? Or just that the cost is not worth optimizing away?
Of course GC is not inherently slower than other memory management strategies. Not only because GC is not a "memory management strategy" but a wide spectrum of them, but also because some GCs are used to speed up memory management compared to low-level languages. Dynamic heap allocations in low-level languages is often minimised because it is slow; some GCs are used to solve this problem.
Have the Rust and Zig communities been outright lying to me? I know very little about the details. I genuinely thought that the overhead of GC was a well-established tradeoff in language design.
That some specific kinds of GCs, specifically moving GCs (of the kind used in Java and .NET, but not the kinds used in Python or Go, which are also very different from each other) can be highly efficient, at least in principle, has been well known since the eighties (see Garbage Collection Can Be Faster Than Stack Allocation, 1986 [1]). However:
1. Moving collectors have long been more efficient than malloc/free but only in the total work devoted to heap management, i.e. throughput. Moving collectors that also offer predictable low latency are much more recent. In fact, the first open-source, production-quality, high-throughput, low-latency moving collector is less than three years old [2].
2. Much more importantly - and this is something that we experienced low-level programmers have known forever but people without much experience in low-level programming seem to not know these days is that low-level languages are not optimised for maximum performance. They're designed for maximum low-level control. When programs are small, when they're not heavily concurrent, or when the hot path is relatively simple, this low-level control can, indeed, translate to very high performance. But when programs grow larger or more concurrent, low-level control can actually make some optimisations harder. In particular, moving pointers, which are required to enjoy the optimisation offered by moving collectors, is not compatible with the low-level control needed in low-level languages, and these languages prioritise low-level control over everything else, including performance (as low-level control is their primary purpose). And this is not the only example of optimisations that these languages make harder. This is why large and/or concurrent programs have largely migrated from C++ to Java and C# over the past few decades, and this trend isn't reversing. Again, for smaller and/or less concurrent programs, low-level languages still offer excellent performance in expert hands, provided you invest sufficient effort into manual optimisation.
3. The optimisation offered by moving collectors isn't free. Until recently, you had to pay in unpredictability and high latency (which is still the case in all languages except Java), and since the algorithm uses RAM to reduce the CPU needed for heap management, it does necessarily require higher footprint. You can enjoy a similar optimisation that turns RAM into free CPU in Zig by using arenas (this is harder to do in C++ and Rust). Zig's arenas are even more efficient than moving collectors, but they do require more effort, and they're less general.
[1]: https://www.cs.princeton.edu/techreports/1986/045.pdf Note that in practice, moving GCs are not quite as efficient as stack allocation (let alone more efficient), but heap allocations in Java are not as expensive as heap allocations in C/C++/Rust/Zig that utilise malloc/free. This is why in these languages we try to avoid heap allocations on the fast path.
The JVM goes for performance, too, only with an emphasis on the performance of larger programs. It's designed to address some of the serious performance issues that large C++ programs tend to suffer from (I originally made the switch to Java because it was getting hard to keep the large C++ applications we were working on fast enough; Java does some optimisations that are hard for a C++ compiler).
One thing we C++ programmers know, though, is that direct control over instructions helps performance in small programs but can hurt it in larger programs. E.g. it is really hard to enjoy the performance benefits that moving collectors and JITs can bring in large C++ applications.
It's not so easy. The pattern you'd want to try and match or beat moving collectors is arenas, and they're really not trivial to use in C++ (or Rust). The only low-level language that supports them well is Zig. But even then, evolution of the program can often require large architectural changes.
So yes, in principle it's possible to match and perhaps somewhat exceed Java's performance in C++, even in large programs (just as, in principle, it's possible to match and exceed C++'s performance in Assembly), but we don't care about what's possible in principle; we care about what we can achieve with the budget we have. Or put another way, C++ has better performance/effort than Assembly, and Java has better performance/effort than C++ in large programs (perhaps not in all domains, but in important ones). The JVM was designed to (among other things) specifically make it easier to overcome some known performance problems that large C++ programs experience.
> we care about what we can achieve with the budget we have
100% agree on this. This is the last driver for everything else in professional environments.
Related: I think C++ safety being driven by profiles and not the Safe C++ stuff that was proposed violates the economic assumption in so many ways that it was the wrong choice for C++.
I think you've misunderstood, because not only would this have been the right design if it had been in Java since day one as it's in line with the philosophy of the platform, it's also simpler than in other languages.
The idea is that instead of controlling memory layout and referencing directly, you communicate your intent: do you care about this value's identity or not? Do you need atomicity or not? Do you need nullability or not? Once the intent is clear, the compiler is free to choose the most appropriate and efficient layout for the particular value at the particular use site. In other words, you say what semantics you're interested in rather than how to implement them at the lowest level.
This opens up optimisation opportunities that are lost when the programmer directly controls the representation rather than the intent. E.g. in other languages you may say whether you want to pass some value by reference or by value. Here the compiler is free to say, well, if identity and nullability are not needed here, I can either pass by value or by reference, and I'll do whichever is more efficient.
And by the way, your point about early misdesign (whether it applies here or not) also inverts the desired state. Every language makes decisions that will be suboptimal in the environment some time later, and Java certainly has its share (its mutability and nullability default; how it treats serialisation). Rust, for example, was first designed twenty years ago, and some of its fundamental decisions reflect the state of the world at that time (it went all-in on some C++ premises that seemed fine 20 years ago). But since important codebases often outlast the outdatendness of early language decisions (your OS and your browser are running some >30yo code and/or affected by >30yo design decisions), one of the things most important in a language isn't the decisions it makes early on - some will prove "wrong" while your codebase is still alive and kicking - but how well it adapts and evolves. So when you pick a language for an important project today, the language's current state will end up mattering less than how the language evolves in the future. The question asked isn't "will I like this decision today?" but "will I regret this decision ten years from now?" Java is one of the languages with the lowest "regret factor", possibly lowest of them all.
Yes, but only in the programming sense of productivity, not in the economic sense. When prices stop being subsidised and the novelty wears off, this will boil down to the question of how much money is that worth? We're already seeing an inflation in such projects, which are becoming hard to even show off. It could be similar to 3D printers. Fun and valuable, with a market size of $25B, but not quite world-changing (I'm not saying AI won't be world-changing, only that this tinkering aspect may not be).
You can, and indeed that's what we JDK developers do.
But even aside from the fact that writing the code is only a very small portion of the effort in this particular project (you can see that the volume of code making its way into the JDK is very small compared to the number of people involved), I'm often amazed by the gap between how well a frontier model (GPT 5.6 Sol in my case) can comprehend code and investigate a bug, and how badly it writes code and documentation, even when it understands things well. So not only is writing the code not a large portion of the effort in this project, it also happens to be one of the things current models don't do as well as other things.
There's a lot of art to using formal methods around how to specify the system at the right level of abstraction (to make verification tractable) and how to specify the correctness properties so they can be easily evaluated. Even with AI assistance as it currently exists, users need to know formal methods well enough to at least understand the specification of the system and the correctness properties, which requires ~90% of the effort of learning formal methods in the world before AI.
But the real hope is that one day AI will be able to use formal methods correctly on its own, benefitting those who don't know formal methods. AI can sometimes do that today, but sometimes isn't good enough for people who don't know formal methods. It is certainly possible that soon enough AI will be able to do this more reliably, but then we get into the hard problem of speculating the "AI future". It is very hard to predict what an AI that can take over the art of using formal methods cannot do. Predicting that AI will be able to do that yet not be able to collect requirements and build software autonomously, or even come up with the idea for what software to build in the first place, or even replace the software's users seems arbitrary to me. In other words, if people think AI will take care of the verification letting us focus on requirement validation, my question would be, why wouldn't an AI that knows how to verify also know how to validate the requirements? For that matter, why wouldn't it also know how to replace the users altogether?
reply