Re: [ANN] LuaJIT-2.0.0-beta1

[Vyacheslav Egorov <mister.aleph@...>]

>
> I need to add auto-vectorization for that. Other than that, it's
> already a killer: someone just sent me a private mail where it
> outperformed Java on heavily numerical code (real use case, not a
> synthetic benchmark).
>

I am curious what was JVM's weak side here (assuming that
measurements/configuration etc are flawless).

Maybe that person will consider sharing his experiences? I am pretty
sure both worlds might benifit from that.

Also there were some benchmarks[1] showing that HotSpot-Server is not
the fastest JVM around. It might be interesting to test the very same
number crunching app on JRockit or Excelsior JET.

PS. BTW, Mike your code is a very insightful reading. Thank you.

[1] http://www.stefankrause.net/wp/?p=9; methodology of the
benchmarking was not flawless, so results should be interpreted with
caution.

--
e.v.e

On Mon, Nov 2, 2009 at 12:01 AM, Mike Pall <mikelu-0911@mike.de> wrote:
> David Given wrote:
>> I assume what you're using here are hand-translated versions of the
>> SciMark benchmarks to each of the various languages? Is it:
>>
>> http://luajit.org/download/scimark.lua
>>
>> ...? What about the C version? I'd be interested to run it through Clue
>> and see what happens.
>
> http://math.nist.gov/scimark2/download.html
>
>> Incidentally, I have tried to compare LuaJIT 2 with V8, but V8 doesn't
>> seem to want to build easily these days. Last time I tried it I recall
>> that it wasn't particularly fast compared to LuaJIT 1 anyway!
>
> http://shootout.alioth.debian.org/u32/benchmark.php?test=all&lang=luajit&lang2=v8&box=1
>
> Looks like V8 is only faster on the only recursive and allocation-
> heavy benchmark in there. And it's much slower on all others.
> Even LJ1 was often faster than V8.
>
> Also check the ranking here:
>  http://shootout.alioth.debian.org/u32/benchmark.php?test=all&lang=all&box=1
>
>  1.00  C++
>  1.02  C
>  1.62  Java steady state
>  1.86  Java
>  2.11  Fortran
>  2.32  ==> LuaJIT <==
>  2.56  C# Mono
>  ....
>  9.77  JavaScript V8
>  ....
> 30.18  Ruby JRuby
> 34.22  JavaScript TraceMonkey
> 43.62  Ruby
> 44.45  Python
> 66.51  Perl
>
> Ok, kicking Fortran will be easy. Next victim: Java.
>
>> Anyway, it's very good to see this out, and it's impressively fast. It'd
>> be very interesting to try doing traditionally high-CPU tasks with
>> LuaJIT 2 and to see how well it works;
>
> I need to add auto-vectorization for that. Other than that, it's
> already a killer: someone just sent me a private mail where it
> outperformed Java on heavily numerical code (real use case, not a
> synthetic benchmark).
>
>> I work in the mobile phone
>> industry and it is *frightening* how shoddy the code tends to be. It'd
>> be really nice to move away from traditional languages like C and C++
>> and towards scripting languages like Lua.
>>
>> Now I just have to wait for an ARM version...
>
> Well, waiting is not a good tactic.
>
> Collect requirements (desired number type, target architectures,
> VFP or not etc.), collect funding, gather people who'd like to
> work on it. In short: let me not drown in administrative stuff.
> I'm more efficient at designing/coding.
>
> [Sorry for the rant, please don't take it personally.]
>
> --Mike
>