What kind of machine are you running this on? Beca...

# advent-of-code

[Joris PZ]

What kind of machine are you running this on? Because my implementation is essentially the same as yours but won't go below ca 40 ms

[joelpedraza]

What are you using to bench?

[Joris PZ]

On the JVM I use measureNanoTime in a loop running my solver 25 times in a row, see https://github.com/jorispz/aoc-2018/blob/master/src/commonMain/kotlin/Runner.kt

[joelpedraza]

Using jmh for mine. I can't fathom what kinds of magic hotspot is doing for it to be faster than native rust

[joelpedraza]

Intel(R) Core(TM) i7-7920HQ CPU @ 3.10GHz

[joelpedraza]

I even tried storing the entire table in the stack and doing unsafe (no bounds checked) access, and it was still slower

[joelpedraza]

I’m guessing some SIMD black magic, just a guess though not at all familiar with JVM internals

[Joris PZ]

Me neither. But in general, the JIT can perform runtime optimizations that a ahead of time compiler just can't do. It's magic

[Joris PZ]

It's very clear in my daily cross-platform measurements, the JVM and JS always get dramatically faster after an iteration or two, while native is always constant across iterations

[joelpedraza]

I always enjoy seeing your cross platform runtime analysis 😁

[Joris PZ]

Thanks! I just checked, I am running on an Intel Core i7-7700HQ @ 2.80 GHz

[Joris PZ]

Oh! It just occurs to me part 2 can easily be run in parallel! Hold my beer!

[Joris PZ]

Running the search for the solution for part 2 in parallel using async brings the best time on the JVM down from 39 to 13 ms

[joelpedraza]

Nice, looking forward to seeing it. I’ve shamefully not yet written any coroutines

[Joris PZ]

Here's a quick version on the JVM. You can replace Dispatchers.Default (which uses all cores) with singleThreadedDispatcher on line 26 to force it to use only one core.

Unfortunately, I haven't been able to get coroutines running on native, so I can't use it in my general repo for now

Foo.kt

[joelpedraza]

Unrelated but i think the difference in perf between your impl an mine is the Triple, which are boxed, new heap allocation everytime a new best is found

[joelpedraza]

Also, summedAreaTable could be inlined

[joelpedraza]

@Joris PZ Got rust down to 4ms

problem11::part_2        ... bench:  10,120,317 ns/iter (+/- 750,638)
problem11::part_2_par    ... bench:   4,053,446 ns/iter (+/- 237,492)

[joelpedraza]

Same technique, data parellization

[Joris PZ]

Wow that's fast, cool!

[joelpedraza]

HOLY WOW. JVM still wins

[joelpedraza]

2.7 ms!

[Joris PZ]

😮

[Joris PZ]

Did you use the coroutine implementation? Or something else?

[joelpedraza]

Java 8 parallel stream

[Joris PZ]

Ah yes

[joelpedraza]

At this point I’m convinced hotpot is memoizing it xD

[joelpedraza]

(sarcasm)