SIMD accelerated sorting in Java – how it works and why it was 3x faster

https://jbaker.io/2022/06/09/vectors-in-java/

by CHY872 • 4 years ago

154 50 4 years ago

geokon3• 4 years ago

Great article. Very comprehensive and well written, while not skipping the gotchas the author hit along the way

I don't really much about how the JVM works, but it's interesting that this is (if I understood correctly) in effect calling out to some native code. In the JVM world, when are things added to the byte code as a new instruction, and when are they wrapped up in a native code call?

I'd expect a new vector instruction for the stack machine, but I'm guessing that's less flexible and less performant..?

"Practically speaking, you end up with code that looks like it creates many objects, but actually compiles to code that allocates no memory. It is extremely disconcerting."

It's interesting comparing people working in Java vs C++. Java makes it so impractical to really reason about the code directly that users ends up relying a lot more on tooling for benchmarking and introspection (and hence the tools are top-class)

"Java’s ability to print its compiler output as assembly "

How do you do this? Is this integrated into the IDE?

While the tools are great and you can dissect all your performance problems, ultimately feedback from the compiler should be feeding back into the editor. I feel that some of the issues the author hit shouldn't happen. You should for instance be immediately able to see if a function is being inlining while you're editing and not having to run extra tools

dzaima4• 4 years ago

You can't really have bytecode instructions for the vector API functions because they're generic. They're instead methods, which the JIT compiler can handle specially if it so chooses to. Effectively, for pure computational operations (i.e. not ones doing complex stack operations or jumps), there's no real difference between a method call and a new bytecode (except that adding a new bytecode is a lot more work because everything must be updated to support it, whereas a new method is just a new method).

pjmlp6• 4 years ago

If you use Oracle's own IDE, it will support it out of the box, as it already did on Sun's days.

Then there are other ways depending on which JVM implementation is used.

On OpenJDK's case you can load runtime plugin to do it

https://github.com/AdoptOpenJDK/jitwatch

geokon3• 4 years ago

Oh very cool. I'm really glad these things exist. I skimmed the intro video and it does looks like a separate tool though. Would be extra slick if these things got highlighted as you coded. But I get that usually you only really bring out the big guns when you're optimizing a tight loop somewhere - and then you probably want the whole toolbox and not just some annotations.

Thanks for letting me know about this :)

pron3• 4 years ago

In the JVM world, when are things added to the byte code as a new instruction, and when are they wrapped up in a native code call?

These aren't native code calls but compiler intrinsics. The Java compiler (the JIT compiler, that is) translates certain Java method invocations into specific machine instruction patterns.

Java makes it so impractical to really reason about the code directly that users ends up relying a lot more on tooling for benchmarking and introspection (and hence the tools are top-class)

As others have mentioned, you can dump the generated machine code, but you are correct that Java is, by design, higher level than C++ in the sense that low level details are less under the programmer's control, but the compiler's. So when you see a `new Foo()` in Java, you cannot tell whether an object will actually be allocated on the heap, or perhaps "scalarized" and stored directly in registers. In fact, what happens is not even decided statically and could vary dynamically, depending on what other code does, as compilation is done JIT, and optimisations are decided based on the program's current execution profile. As a result, even though all instance methods in Java are semantically virtual, in practice they are very often inlined, and don't compile even to a call instruction, let alone a vtable call.

Unlike C++, Java's goal — generally speaking — is to get very good average-case performance for little effort rather than give the programmer a lot of knobs to control the worst case. Of course, sometimes there are knobs — just like the Vector API, which is useful in cases the JIT cannot automatically vectorise loops — but the balance is different from C++.

Const-me4• 4 years ago

Vector API, which is useful in cases the JIT cannot automatically vectorise loops

For vertical-only operations (vectorizable loops), people often using GPUs instead of SIMD in CPUs. Graphics cards have way more resources, both GFlops and memory bandwidth.

I don't think sorting an array has anything to do with vectorizing a loop.

geokon3• 4 years ago

Thanks for the explanation. So I'm gathering that the distinction is inlined native code will be explicit while an intrinsic is up to the compiler. The intrinsic provides a fallback if you run on a dumb/old JVM. That does seem much cleaner than adding new byte codes

chrisseaton• 4 years ago

You should for instance be immediately able to see if a function is being inlining while you're editing

But it’s a dynamic decision, except in the most trivial cases. How can you know it without running the program?

kaba02• 4 years ago

Would some of the mentioned drawbacks be solved by having Project Panama's other half be available? What I mean is allocating memory manually, instead of using Java primitive arrays inside the Vector API. Would the (de)serialization overhead be too much for it to be worth it?

grishka2• 4 years ago

This whole "the JVM inlines things for you and avoids the allocation of objects altogether sometimes" goes so against my intuition after more than a decade of writing Java, mostly for Android. The idea that you should allocate as few short-lived objects as possible has stuck with me. So these new APIs do make me feel uncomfortable a bit, they make me think "oh man, this does look like it's gonna have some performance issues due to too much time spent in GC deleting all those objects".

I do understand that the regular (Hotspot?) JVM must be much smarter with optimizations than the Android runtime, but still.

kaba02• 4 years ago

Android’s Java is pretty much a bastard child of OpenJDK - it is so behind on features that it is truly a shame.

my123• 4 years ago

Android 13 started to give some OpenJDK 11 features... but ugh - nothing that we can widely rely on yet.

Core Java 11 features will be backported to Android 12 with ART as an APEX module, but that didn't happen quite yet as far as I can see.

dale_glass2• 4 years ago

I think this problem occurs at many places in the industry. Developer learns a thing, then keeps trying to apply it two decades later, long after it became completely obsolete.

That's how you get things like people unrolling a loop by hand. Compilers have known how to do it for ages, and have built-in data about what works best on what CPU. 99% of the time they do a better job than a programmer that developed a pattern back when they were writing for a 386.

These days it often works better to write very straightforward, non-clever code and let the compiler figure out how to optimize it.

Const-me4• 4 years ago

Here’s an example of a simple C++ function, manually vectorized + unrolled for optimal performance on modern processors: https://stackoverflow.com/a/59495197/126995

Of all C++ compilers, only clang generates similar machine code from a straightforward source. And only with some special compiler switches, most importantly the -fast-math. Dot product of 2 vectors is a trivially simple algorithm, sum( a[ i ] * b[ i ], i = 0..N-1 ), I wouldn’t expect clang to auto-vectorize more complicated stuff. Finally, C++ compilers are designed to work offline so the optimizer can afford to spend substantial CPU time searching for a best optimization, a JIT compiler like Java runtime simply doesn’t have time for that.

pjmlp6• 4 years ago

Android Java is not like proper Java, including how ART works. At least they dropped Dalvik which was comparable to early JVM implementations pre-JIT days, think Java 1.2.

Then there is the whole issue that Google isn't that up the game into making the same kind of optimizations as the Java community, and rather calls into C++ via native methods than optimizing ART more than good enough.

Finally the sore state of proper Java support comes from their beef with Sun and Oracle initially, and now their agenda to push Kotlin no matter what, and every modern Java feature that becomes one red less against Kotlin doesn't help.

Hence why most of their Java vs Kotlin samples are mostly based on Java 7.

If the Android team really cared, they would provide equal support for both languages and then let the community pick up which one they were found of.

Even the recently announced improved support for Android 13 is based on a Java 11 LTS subset, when the more recent LTS version is 17, and Java 19 is a couple of months away.

vips7L2• 4 years ago

Fwiw, as a Java fanboy, I do think Kotlin/Compose is way better for creating GUI’s than what Java currently has to offer. I recently went and tried some swing again and it wasn’t enjoyable, but like you said it’s like this for a reason.

layer84• 4 years ago

The idea that you should allocate as few short-lived objects as possible has stuck with me.

I don’t know about Android, but since generational garbage collectors have become the default, the rule is that short-lived objects are dirt cheap, because they’ll be collected all at once in O(1) with the young generation. And since allocating them is usually just a pointer increment, they can be as cheap as stack objects.

snek_case• 4 years ago

since allocating them is usually just a pointer bump, they can be as cheap as stack objects.

I don't think that's quite true. Even with a copying/moving GC, you still need to traverse all the live objects and then copy all of them. Asymptotically as cheap as stack objects maybe, but in reality the overhead is greater. GCs also have some amount of overhead due to all the synchronization they need to do.

dc-programmer4• 4 years ago

I’m curious how the JVM efficiently allocates many small objects. How does it avoid memory fragmentation? Or does it request a large block of memory from the OS at the start? (last q is probably easily googled)

kwizzt• 4 years ago

I'm really looking forward to when SIMD support is released in later versions of Java. If anyone here follows handmade hero (by Casey), he explained SIMD in some of his videos when implementing the software renderer. The speedup from regular code to SIMD is very impressive. You essentially get speedup of 2x-3x or even more for free.

The downside is it makes the code much harder to read, so I think it should be applied only in very small sections of the code.

ErikCorry• 4 years ago

The section on "dictionary encoding" didn't make much sense to me. Is there an explanation of this technique somewhere? The example code took a list called "buffer" and then didn't use it for anything - perhaps something got lost when writing the code?

The explanation was pretty opaque: Here’s one simple example I encountered recently. It’s common in columnar database code to ‘dictionary encode’ lists of values. Here, for some subset of the stored data, we store a dictionary containing the unique values that exist, assigning unique indices to each value, and we store a list of pointers into the list to represent each element.

We store a list of pointers into which list? The dictionary contains the unique values as keys or as values? How do we pick which subset of the stored data we do this for?

kwillets• 4 years ago

The buffer in this case is the input to a max() function; his point is simply that sorting the dictionary indices before looking them up is a good idea for common data structures like arrays, disk blocks, or anything with caching or readahead.

SIMD accelerated sorting in Java – how it works and why it was 3x faster

Related Stories