Wild performance tricks(davidlattimore.github.io) |
Wild performance tricks(davidlattimore.github.io) |
I frequently encounter use-cases akin to the “Sharded Vec Writer” idea, and I agree it can be valuable. But if performance is a genuine requirement, the implementation needs to be very different. I once attempted to build a general-purpose trait for performing parallel in-place updates of a Vec<T>, and found it extremely difficult to express cleanly in Rust without degenerating into unsafe or brittle abstractions.
To say more about it: nearly any modern high performance allocator will maintain a local (private) cache of freed chunks.
This is useful, for example, if you're allocating and deallocating about the same amount of memory/chunk size over and over again since it means you can avoid entering the global part of the allocator (which generally requires locking, etc.).
If you make an allocation while the cache is empty, you have to go to the global allocator to refill your cache (usually with several chunks). Similarly, if you free and find your local cache is full, you will need to return some memory to the global allocator (usually you drain several chunks from your cache at once so that you don't hit this condition constantly).
If you are almost always allocating on one thread and deallocating on another, you end up increasing contention in the allocator as you will (likely) end up filling/draining from the global allocator far more often than if you kept in on just one CPU. Depending on your specific application, maybe this performance loss is inconsequential compared to the value of not having to actually call free on some critical path, but it's a choice you should think carefully about and profile for.
So in C++/WinRT, which is basically the current C++ projection for COM and WinRT components, the framework moves the objects into a background thread before deletion, as such that those issues don't affect the performance of the main execution thread.
And given it is done by the same team, I would bet Rust/Windows-rs has the same optimization in place for COM/WinRT components.
That's weird. I'd expect it to work with _any_ type, primitive or not, newtype or not, with a sufficiently simple memory layout, the rough equivalent of what C++ calls a "standard-layout type" or (formerly) a "POD".
I don't like magical stdlibs and I don't like user types being less powerful than built-in ones.
Clever workaround doing a no-op transformation of the whole vector though! Very nearly zero-cost.
> It would be possible to ensure that the proper Vec was restored for use-cases where that was important, however it would add extra complexity and might be enough to convince me that it’d be better to just use transmute.
Great example of Rust being built such that you have to deal with error returns and think about C++-style exception safety.
> The optimisation in the Rust standard library that allows reuse of the heap allocation will only actually work if the size and alignment of T and U are the same
Shouldn't it work when T and U are the same size and T has stricter alignment requirements than U but not exactly the same alignment? In this situation, any U would be properly aligned because T is even more aligned.
// This compiles down to a memmove() call
let my_vec: Vec<_> = my_vec.into_iter().skip(n).collect();
// this results in significantly smaller machine code than `v.retain(f)`
v.into_iter().filter(f).collect();
For non-performance-sensitive code, sure, go ahead and rely on the rust compiler to compile away the allocation of a whole new vector of a different type to convert from T to AtomicT, but where the performance matters, for my money I would go with the transmute 100% of the time (assuming the underlying type was decorated with #[transparent], though it would be nice if we could statically assert that). It'll perform better in debug mode, it's obvious what you are doing, it's guaranteed not to break in a minor rustc update, and it'll work with &mut [T] instead of an owned Vec<T> (which is a big one).
Though this optimisation is treated as an implementation detail [1].
[1]: https://doc.rust-lang.org/stable/std/vec/struct.Vec.html#imp...
I liked most of the tricks but this one seems pointless. This is no different than transmute as accessing the borrower requires an assume_init which I believe is technically UB when called on an uninit. Unless the point is that you’re going to be working with Owned but want to just transmute the Vec safely.
Overall I like the into_iter/collect trick to avoid unsafe. It was also most of the article, just various ways to apply this trick in different scenarios. Very neat!
Those optimisations that this code relies on are literally undefined behaviour. The compiler doesn't guarantee it's gonna apply those optimisations. So your code might suddenly become super slow and you'll have to go digging in to see why. Is this undefined behaviour better than just having an unsafe block? I'm not so sure. The unsafe code will be easier to read and you won't need any comments or a blog to explain why we're doing voodoo stuff because the logic of the code will explain its intentions.
This might be related in part to the fact that Rust chose to create specific AtomicU8/AtomicU16/etc. types instead of going for Atomic<T> like in C++. The reasoning for forgoing the latter is [0]:
> However the consensus was that having unsupported atomic types either fail at monomorphization time or fall back to lock-based implementations was undesirable.
That doesn't mean that one couldn't hypothetically try to write from_mut_slice<T> where T is a transparent newtype over one of the supported atomics, but I'm not sure whether that function signature is expressible at the moment. Maybe if/when safe transmutes land, since from_mut_slice is basically just doing a transmute?
> Shouldn't it work when T and U are the same size and T has stricter alignment requirements than U but not exactly the same alignment? In this situation, any U would be properly aligned because T is even more aligned.
I think this optimization does what you say? A quick skim of the source code [1] seems to show that the alignments don't have to exactly match:
//! # Layout constraints
//! <snip>
//! Alignments of `T` must be the same or larger than `U`. Since alignments are always a power
//! of two _larger_ implies _is a multiple of_.
const fn in_place_collectible<DEST, SRC>(
step_merge: Option<NonZeroUsize>,
step_expand: Option<NonZeroUsize>,
) -> bool {
if const { SRC::IS_ZST || DEST::IS_ZST || mem::align_of::<SRC>() < mem::align_of::<DEST>() } {
return false;
}
// Other code that deals with non-alignment conditions
}
[1]: https://github.com/rust-lang/rust/blob/c58a5da7d48ff3887afe4...
Cool. Thanks for checking! I guess the article should be tweaked a bit --- it states that the alignment has to match exactly.
Not really. Panics are supposed to be used in super exceptional situations, where the only course of action is to abort the whole unit of work you're doing and throw away all the resources. However you do have to be careful in critical code because things like integer overflow can also raise a panic.
So you can basically panic anywhere. I understand people have looked at no-panic markers (like C++ noexcept) but the proposals haven't gone anywhere. Consequently, you need to maintain the basic exception safety guarantee [1] at all times. In safe Rust, the compiler enforces this level of safety in most cases on its own, but there are situations in which you can temporarily violate program invariants and panic before being able to restore them. (A classic example is debiting from one bank account before crediting to another. If you panic in the middle, the money is lost.)
If you want that bank code to be robust against panics, you need to use something like https://docs.rs/scopeguard/latest/scopeguard/
In unsafe Rust, you basically have the same burden of exception safety that C++ creates, except your job as an unsafe Rust programmer is harder than a C++ programmer's because Rust doesn't have a noexcept. Without noexcept, it's hard to reason about which calls can panic and which can't, so it's hard to make bulletproof cleanup paths.
Most Rust programmers don't think much about panics, so I assume most Rust programs are full of latent bugs of this sort. That's why I usually recommend panic=abort.
[1] https://en.wikipedia.org/wiki/Exception_safety#Classificatio...
This is incorrect. Only in debug builds does it raise a panic. In release Rust has to make the performance tradeoff that C++ does and defines signed integer math to wrap 2’s complement. Only in debug will signed overflow panic. Unsigned math never panics - it’s always going to overflow 2’s complement.
For example, in something like Go (which has a weaker type system than Rust), you wouldn't think twice about, paying for the re-allocation in buffer-reuse example.
Of course, in something like C or C++ you could do these things via simple pointer casts, but then you run the risk of violating some undefined behaviour.
No you don't. You explicitly start a new object lifetime at the address, either of the same type or a different type. There are standard mechanisms for this.
Developers that can't be bothered to do things correctly is why languages like Rust exist.
People using systems languages more often than not go down the rabbit hole of performance tuning, many times without a profiler, because still isn't the amount of ms that is supposed to be.
In reality unless one is writing an OS component, rendering engine, some kind of real time constrained code, or server code for "Webscale", the performance is more than enough for 99% of the use cases, in any modern compiler.
You cannot get undefined behavior in Rust without an unsafe block.
> The compiler doesn't guarantee it's gonna apply those optimisations.
This is a different concept than UB.
However, for the "heap allocation can be re-used", Rust does talk about this: https://doc.rust-lang.org/stable/std/vec/struct.Vec.html#imp...
It cannot guarantee it for arbitrary iterators, but the map().collect() re-use is well known, and the machinery is there to do this, so while other implementations may not, rustc always will.
Basically, it is implementation-defined behavior. (If it were C/C++ it would be 'unspecified behavior' because rustc does not document exactly when it does this, but this is a very fine nitpick and not language Rust currently uses, though I'd argue it should.)
> So your code might suddenly become super slow and you'll have to go digging in to see why.
That's why wild has performance tests, to ensure that if a change breaks rustc's ability to optimize, it'll be noticed, and therefore fixed.
But benchmarks won't tell us which optimisation suddenly stopped working. This looks so similar to the argument against UB to me. Something breaks, but you don't know what, where, and why.
Correctness in debug builds is important, isn't it?
That said, panic on integer overflow in debug builds is unfortunate behavior. Overflow should cause an abort, not a panic.
> make the performance tradeoff that C++ does and defines signed integer math to wrap 2’s complement
In C++, signed overflow is undefined behavior, not wraparound. This property is useful to the optimizer for things like inferring loop bounds. The optimizer has less flexibility in equivalent Rust code.
Personally I would often choose both, overflow panics and also panics abort, so if we overflow we blow up immediately.
Software is built on abstractions - if all your app code is written without unsafe and you have one low level unsafe block to allow for something, you get the value of rust for all your app logic and you know the actual bug is in the unsafe code
The point is to escalate capability only when you need it, and you think carefully about it when you do. This prevents accidental mistakes having catastrophic outcomes everywhere else.
It’s a speed bump that makes you pause to think, and tells reviewers to look extra closely. It also gives you a clear boundary to reason about: it must be impossible for safe callers to trigger UB in your unsafe code.
Also the community culture matters, even though static analysis exists for C since 1979, it is still something we need to force feed many developers on C and C++ world.
Foo foo{}; init(*(Bar *)foo);
For implicit-lifetimes types (iirc types with trivial default constructors (or are aggregates) plus trivial destructors), you can use memcpy, bit_cast and soon std::start_lifetime_as (to get a pointer) when it is implemented.
If I'm not mistaken, in C, the lifetime rules are more or less equivalent to implicitly using C++'s start_lifetime_as
UB is really about the observable behavior of the abstract machine which is limited to the reads/writes to volatile data and I/O library calls [1]
[1] http://open-std.org/jtc1/sc22/open/n2356/intro.html
Edit: to clarify the example
You mentioned particularly the C++ unstable sort std::sort. Famously although C++ 11 finally guarantees O(n log n) worst case complexity the libc++ stdlib didn't conform. They'd shipped worst case O(n squared) instead.
The bug report saying essentially "Hey, your sort is defective", was opened in 2014. By Orson Peters. It took until 2021 to fix it.
You can easily write code that does not contain any possible panic points, if you want.
For example: do logging frameworks guarantee no-panic behavior? People can add logging statements practically anywhere, especially in a large team that maintains a codebase over significant time. One innocuous-looking debug log added to a section of code that's temporarily violated invariants can end up putting the whole program into a state, post-panic, in which those invariants no longer hold.
A lot of experience tells us that this happens in practice in C++, Java, Python, and other excpeption-ful languages. Maybe it happens less in Rust, but I'd be shocked if this class of bugs were absent.
Note that I'm talking about both safe and unsafe code. A safe section of code that panics unexpectedly might preserve memory safety invariants but hork the higher-level logical invariants of your application. You can end up with security vulnerabilities this way too.
Imagine an attacker who can force a panic in a network service, aborting his request but not killing the server, such that the panic on his next request grants him some kind of access he shouldn't have had due to the panic leaving the program in a bad state.
I'm not seeing Rust people take this problem as seriously as I think is warranted.
The usual way of dealing with this is to use impl Drop to cleanup properly. Resources are guaranteed to be dropped as expected on panic unwinds. Eg the database transaction rolls back if dropped without committing.
> Imagine an attacker who can force a panic in a network service, aborting his request but not killing the server, such that the panic on his next request grants him some kind of access he shouldn't have had due to the panic leaving the program in a bad state.
You need to be more specific. Why would the web server be left in a bad state because of such panics (in safe rust). All the memory will be cleaned up, all the database transactions will be cleaned up, mutexes might get poisoned, but that's considered a bug and it'll just cause another panic the next time someone tries to lock the mutex.
Wouldn't a relational database help deal with this?
> Without noexcept, it's hard to reason about which calls can panic and which can't, so it's hard to make bulletproof cleanup paths.
Unsafe blocks usually contain very low level code, so you can understand what your code does very accurately. If the unsafe code calls a dependency which calls 150 other dependencies transitively, yeah, that's going to be pretty bad.
Sure. It's just a toy example. There are lots of real programs in which you temporarily violate invariants in the course of performing some task, then restore them after. Another example that comes to mind is search tree rotations.
> Unsafe blocks usually contain very low level code, so you can understand what your code does very accurately.
Perhaps at first, but code evolves.
I'm doubtful that those boundaries that you mention really work so great. I imagine that in practice you can easily trigger faulty behaviours in unsafe code from within safe code. Practical type systems are barely powerful enough to let you inject a proof of valid-state into the unsafe-call. Making a contract at the safe/unsafe boundary statically enforceable (I'm not doubting people do manage to do it in practice but...) probably requires a mountain of unessential complexity and/or runtime checks and less than optimal algorithms & data structures.
We agree that this is a dangerous / security-defeating habit to develop.
If someone realizes they're developing a pattern of such commands, it might be worth considering if there's an alternative. Some configuration or other suid binary which, being more specialized or tailor-purpouse, might be able to accomplish the same task with lower risk than a generalized sudo command.
This is often a difficult task.
Some orgs introduce additional hurdles to sudo/admin access (especially to e.g. production machines) in part to break such habits and encourage developing such alternatives.
> unsafe
There are usually safe alternatives.
If you use linters which require you to write safety documentation every time you break out an `unsafe { ... }` block, and require documentation of preconditions every time you write a new `unsafe fn`, and you have coworkers who will insist on a proper soliloquy of justification every time you touch either?
The difficult task won't be writing the safe alternative, it will be writing the unsafe one. And perhaps that difficulty will sometimes be justified, but it's not nearly so habit forming.
And strict aliasing is not a concern due to Rust's aliasing models, thus the combination of the two makes it safe to type-pun like that. But Rust's models has its downsides/is a tradeoff, so...
I don't particularly mind the C++ object model (since C++20), it makes sense after all: construct your object if it needs to, or materalize it through memcpy/bit_cast. std::start_lifetime_as should fix the last remaining usability issue about the model.
Sure, though that's because C has abstraction like Mars has a breathable atmosphere.
> This approach leads to straightforward, efficient architecture and bug-free code. It's also much better for concurrency/parallelism.
This claim is wild considering that Rust code is more bug-free than C code while being just as efficient, while keeping in mind that Rust makes parallelism so much easier than C that it's stops being funny and starts being tragic.
Plenty of abstraction possible using TU as modules, and applying Abstract Data Types design, while following Yourdon structured method with C.
The grace with which C handles projects of high complexity disagrees.
You get a simple implementation only by ignoring edge cases or improvements that increase complexity.
I'd say, more fully featured languages are most useful for the simpler side of projects (granted some of them can scale quite a way up with proficient use).
Now go research how some of the most complex, flexible, and efficient pieces of software are written.
I think this is wrong on its face. We wouldn't see any correlation between the language used and the highest complexity programs achieved it in.
As recently mentioned on HN it takes huge amounts of assembly to achieve anything at all, and to say that C doesn't handle any of the complexity you have to deal with when writing assembly to achieve the same result is absurd.
EDIT: > Now go research how some of the most complex, flexible, and efficient pieces of software are written.
I'm quite aware. To say that the choice of say, C++ in the LLVM or Chromium codebase doesn't help deal with the complexities they operate over, and that C would do just as well at their scale... well, I don't think history bears that out.
I'm not sure that LLVM would be the first consideration for complex, flexible, efficient? It's quite certainly not fast, in particular linking isn't. I'm not sure about Chromium, it would be interesting to look at some of the more interesting components like V8, rendering engine, OS interfacing, the multimedia stack... and how they're actually written. I'd suspect the code isn't slinging shared_ptr's and unique_ptrs and lambdas and is keeping use of templates minimal.
I would have thought of the Linux kernel first and foremost. It's a truly massive architecture, built by a huge number of developers in a distributed fashion, with many intricate and highly optimized parts, impressive concurrency, scaling from very small machines to the biggest machines on the planet.
This is what the Linux kernel achieved, and when it started C was definitely the right choice for its (primary) implementation language. That doesn't mean C made the job easier, or that if Linux started today it would follow the same trajectory.
I would say Linux succeeds and has good abstractions due to strong discipline, in spite of what C provides it.
But sure, if you’re that confident go for it. Writing a linker that can use a fraction of the memory, especially during LTO would be a ground breaking achievement.
For sure, linking "graphs" generally have random connections so obviously I'm not saying that linking in general is a trivial stream processing problem.
Thinking about it though, linking graphs are _almost_ free of cyclic dependencies in practice (even when looking at entire chunks of code or whole object files as units), so there are some basic tricks we can use so we don't need to load all compilation units into RAM as I think you claimed we need to. I don't think it's necessary to load all the data in RAM.
The way I understand it, linking is essentially concatenating object files (or rather the sections contained in them) into a final executable. But while doing so, relocations must be applied, so object code must not be written before all symbols referenced by that code are known. This can be done in a granular fashion, based on fixed-size blocks. You collect all relocations in per-block lists (unsorted, think std::vector but it can be optimized). When a symbol or code location becomes known (i.e. looking at the library that has the required symbol), that information gets appended to the list. When a list is full (i.e. all relocations known) that block can be committed to the file while applying all fixups. Applying fixups is still random writes but only to fixed size blocks.
That way the problem is almost reduced to streaming fixed size blocks in practice. Typically, all symbols that some object file depends on are typically resolved by the object files that were already added before. So most new chunks that we're looking at can immediately be fixed up and be streamed to the output file. (In fact some linkers are kind of strict about the order of objects that you pass on the command line). What does grow more-or-less linearly though is the set of accumulated symbols (e.g., name + offset) as we add more and more objects files to the output executable.
I don't know a lot about LTO but it seems just like an extension of the same problem. Balancing memory consumption and link times with output performance is partly the user's job. Most projects should probably just disable LTO even for release builds. But making LTO fast in a linker is probably all about using the right approach, doing things in the right order, grouping work to avoid reads and writes (both RAM and disk).
Sort of yes, sort of no. Yes in that the main alternative (std::mem::transmute) is unsafe, so if you want to avoid dealing with that (e.g., "I personally like the challenge of doing things without unsafe if I can, especially if I can do so without loss of performance") then this can indeed be described as a language-inflicted problem. No in that this is arguably not an "inherent" language-inflicted problem, and there is work towards ways for the Rust compiler to guarantee that certain transmutes will work as expected in safe code, and I want to say that this would be one of those cases.