Hacker News Reader

Comment by marginalia_nu | Hacker News Reader

I think in general these types of benchmarks makes boxed numbers look bad only because they're being compared to adding integers, which is a ridiculously fast operation to start with. Accomplishing half that speed despite "dereferencing a pointer" is kinda insane in the light of 'Latency Numbers Every Programmer Should Know'.

(The reason it works is because Java doesn't actually allocate new Longs for small numbers, it fetches them from a table of constants; it's always the same 256 objects that are being dereferenced. I don't know their memory layout, but I'd half expect them to be sequential in memory as that's would be much a low hanging fruit optimization. Optional<Long>'s performance is what you'd expect without these optimizations. Also in this scenario you really should use OptionalLong instead of Optional<Long> but that's beside the point ;-)

View on HN

I think the real issue is that you probably should worry about this overhead only in the context of really tight loops where you are basically wasting a lot of memory allocation and garbage collection for no good reason. But otherwise, Java is actually pretty good dealing with short lived small objects. The garbage collector is pretty good for that.

Normal usage of this stuff is not going to cause any more issues than trigger the OCD of people that obsess about this stuff. And in the rare case that you do have an issue, you can do something more optimal indeed.

If you're writing a database engine, and you accidentally wrap each table cell in a Java object, you'll very quickly hit a GC performance wall, particularly when using a low pause GC. So we try to write Java without objects. Guess how easy that is. ;)

Yeah I deal with that a lot with my search engine index too. Honestly it's not that bad once you get used to it.

You can get away with referencing the data through (mutable and reusable) pointer objects that reference memory mapped areas yet provide a relatively comfortable higher level interface. This gets rid of object churn while keeping a relatively sane interface.

Can you give an example how to do that? I've only seen IO and parsing libraries that copy around byte arrays, but you can't cast byte arrays to arbitrary objects, which means you usually have to manipulate the bytes directly (unless you're lucky to work with only strings). At that point I imagine it would be far easier to just use C++ or Rust.

Yeah this sort of work isn't Java's strong suite. A lot of it's sort of like programming old-school C with oven mitts. It'll get a bit better with the Foreign Memory API which is in the JEP pipe.

But a very bare bones example might look something like

  class MdrLayout {
     static final int FOO_OFFSET = 0;
     static final int BAR_OFFSET = 4;
     static final int BAZ_OFFSET = 12;
     static final int ENTRY_SIZE = 13;
  }

  class MyDataRecord {

    int idx = Integer.MIN_VALUE;
    ByteBuffer buffer = null;

    MyDataRecord() { }

    void movePointer(ByteBuffer buffer, int idx) { 
      this.buffer = buffer;
      this.idx = idx; 
    }

    int foo() { return buffer.getInt(ENTRY_SIZE * idx + FOO_OFFSET); }
    long bar() { return buffer.getLong(ENTRY_SIZE * idx + BAR_OFFSET); }
    byte baz() { return buffer.get(ENTRY_SIZE * idx + BAZ_OFFSET); }

    // may have operators like reset(), next() or skip() as well
  }

  class MyDataRecord2 {

    final ByteBuffer buffer;

    MyDataRecord2(ByteBuffer buffer) { this.buffer = buffer; }

    int foo(int idx) { return buffer.getInt(ENTRY_SIZE * idx + FOO_OFFSET); }
    long bar(int idx) { return buffer.getLong(ENTRY_SIZE * idx + BAR_OFFSET); }
    byte baz(int idx) { return buffer.get(ENTRY_SIZE * idx + BAZ_OFFSET); }
  }

Yes, this is the way. Surprising how high level you can make this work - I’ve seen Collection implementations and all sorts.

If you do stuff like that, use a profiler and identify and fix your real performance bottlenecks. As opposed to applying premature optimization blindly. Same with GC tuning. This has gotten easier over the years. But there are still lots of tradeoffs here.

There are plenty of fast performing databases and other middleware written in Java. The JVM is a popular platform for that kind of thing for a good reason. Writing good software of course is a bit of a skill. Benchmarks like this are kind of pointless. Doing an expensive thing in a loop is slow. Well duh. Don't do that.

GC performance isn't the only thing you need to care about here. Iterating over an ArrayList<T> in Java already makes you chase pointers all over the place due to Java's lack of value types, which wrecks cache locality. ArrayList<Optional<T>> just makes you chase two pointers for each object instead.

List<Optional<T>> is pure insanity fwiw. Filter out your empties before making the list.

Sure, but Optional<T> is what we were already discussing, and List<T> is the collection type that most obviously requires no indirection. Feel free to replace Optional<T> with any other sort of sum type and the point still stands — you always have one extra object between the list and the "actual value".

> List<T> is the collection type that most obviously requires no indirection.

Collections in Java as of now (before project Valhalla) are actually collections of pointers to objects, so at least one level of indirection. Another level of indirection is List itself, which is an interface, so the compiler can’t ahead of time devirtualise method calls, only after JIT detects which List implementation it’s dealing with at runtime.

Virtual calls are trivial optimizations and doesn’t really matter here either way as it only has to happen once, no matter the list size.

Sorry, I wasn't clear: Tree maps have built-in indirection, hash maps can use open addressing but they can also use chaining, which has indirection. An array-backed list is the only collection type that obviously requires no indirection in concept. In a world where Java has value types, you could build ArrayList<T> with no indirection (other than polymorphism/dispatch).

Nothing insane about it, e.g. you might be using some sort of SoA structure and that specific component is optional.

If you're using an SoA pattern you're already optimizing for performance and, imo, should be using null to begin with.

That's the point, though. Java's Optional<T> always incurs that extra cost, but e.g. Kotlin's T? doesn't have that (because ? just allows null), and Rust has an optimisation for Option<Box<T>> that reduces the representation to a single pointer (and the empty case is a null pointer). You absolutely can have that level of optimisation while still having the optional type.

What if you have an in-memory DataFrame? Not hard to end up with something that looks like List<Optional<T>> (say, NullableSeries<T>) as one of the columns.

Java uses thread local allocation buffers such that objects allocated one after another are typically contiguous in RAM. Most modern Java gcs are also compacting, meaning that the heap ends up approximately breadth first ordered after GC.

What this means is that in practice, pointer chasing is less of an issue than you’d expect. Even a linked list will end up with decent cache locality.

Obviously this won’t always work, but it generally works a lot better than the same structure in a systems language.

https://shipilev.net/jvm/anatomy-quarks/11-moving-gc-localit...

The table of constants leads to this fun example:

System.out.println(Integer.valueOf(22) == Integer.valueOf(22)); // true

System.out.println(Integer.valueOf(2200) == Integer.valueOf(2200)); // false

Which is a bit confusing to say the least. I realize one should never be using == with objects, but still.

That's another quality of life improvement in Kotlin, == calls the equals method on the objects, not comparing their references. Which is what most java programmers want in 99% of cases. And makes the code more readable, a.equals(b) is harder to read than a == b.

In most cases I'd use equals(a, b) in Java (with a static import of Objects.equals). It's both null safe and more readable.

I kind of like that behaviour. You're comparing object references and get exactly the expected output instead of some magically overloaded equality operator. If you want that, just use 'equals' ;)

Similar thing happens in Python (assuming my knowledge hasn't gone out of date)

  x = 2
  print(x is 2) // true
  x = 200
  print(x is 200) // false

I think a big difference is Java you have to switch between `==` and `Object.equals` depending whether your type is a primitive (e.g. int) or a reference (e.g. Integer).

In Python the cases where you’d be using `is` are a lot more restricted, and a bit of an optimisation (although most style guides will require it in the cases where it makes sense).

There’s basically 3 cases where you’d use `is` in Python:

- the standard singletons None, True, and False (and for the latter two you’d usually use truthiness anyway)

- placeholder singletons (e.g. for default values when you need something other than None)

- actual identity check between arbitrary objects (which very rarely happens IME)

This will also be “fixed” once Valhalla arrives.

The cached range is -128 to 127. I just reran these benchmarks with ints instead of longs and the mask changed to 127(0x7F) and sumNulls and sumSimple results are exactly the same: 0.6 ns/op. As for the sumOptional method, changing Optional<Integer> for the primitive variant OptionalInt, doesn't change the result much, its the actual creation of the Optional object itself that dominates the time.

In nutshell, on my old i5 2500k:

ints 0.6ns/op

cached Integer 0.6ns/op

boxed Integer 1.3ns/op

OptionalInt 3.5ns/op

Optional<Integer> 4.2ns/op (time includes boxing the int)

Where an op is getting the number, checking it, then an addition.

For hot loops inside a jmh benchmark, you can use @OperationsPerInvocation(MAX) and it will spit out the results in this more readable format for the time just inside the loop.

Benchmark structure is one thing I hope more languages copy from Go - hard-coding iterations is pretty silly, doomed to need repeated changes or become problematically imprecise as hardware and runtime changes occur.

Moreso, non-highly specialised use cases for optionals in this class of language are mostly commonly used around IO ops of some kind (DB, streams, messaging, API etc) so we're into "make sure there's no flies on the elephant when we weight it" territory.

Reply on news.ycombinator.com