Comentarios en: Lua JIT (Just in Time Compiler) http://www.pplux.com/2005/12/06/lua-jit-just-in-time-compiler/ el blog de PpluX Mon, 01 Dec 2008 21:04:00 +0000 http://wordpress.org/?v=2.6.2 By: Mike http://www.pplux.com/2005/12/06/lua-jit-just-in-time-compiler/#comment-64 Mike Tue, 06 Dec 2005 22:35:24 +0000 http://www.pplux.com/?p=55#comment-64 Yes, the Lua interpreter is already very fast. But this is only part of the reason why the LuaJIT speedups are 'conservative'. The published speedups come from 'real' benchmarks and not from some micro benchmarks (usually one-liners). I can easily come up with several micro 'benchmarks' that show a 30x speedup with LuaJIT. But I regard this as dishonest. Telling people they won't see a speedup if they mainly use C library calls (i.e. the 1.04x figure) is maybe bad marketing, but also something they need to consider. For a real cross-language comparison look at http://shootout.alioth.debian.org/ and at http://shootout.alioth.debian.org/gp4/ . Please note that Lua gets its speed without using tons of calls to C functions (like Python does) and still beats Python and Perl easily. Then compare Lua to Psyco (only in the gp4 set) and you'll be enlightened. LuaJIT is not in there (yet), but would score significantly higher on the pure CPU benchmark score (change the weights to 1, 0, 0 in the form). Somewhere around the Java scores (JIT compiled Java, not the Java interpreter). E.g. LuaJIT 1.0.3 gets around 50% of the speed of fully-optimized and hand-tuned C code in the mandelbrot benchmark. That's not so bad for a JIT engine that fits in only 25 Kilobytes IMHO. Yes, the Lua interpreter is already very fast. But this is only part of the reason why the LuaJIT speedups are ‘conservative’.

The published speedups come from ‘real’ benchmarks and not from some micro benchmarks (usually one-liners). I can easily come up with several micro ‘benchmarks’ that show a 30x speedup with LuaJIT. But I regard this as dishonest.

Telling people they won’t see a speedup if they mainly use C library calls (i.e. the 1.04x figure) is maybe bad marketing, but also something they need to consider.

For a real cross-language comparison look at http://shootout.alioth.debian.org/ and at http://shootout.alioth.debian.org/gp4/ . Please note that Lua gets its speed without using tons of calls to C functions (like Python does) and still beats Python and Perl easily. Then compare Lua to Psyco (only in the gp4 set) and you’ll be enlightened.

LuaJIT is not in there (yet), but would score significantly higher on the pure CPU benchmark score (change the weights to 1, 0, 0 in the form). Somewhere around the Java scores (JIT compiled Java, not the Java interpreter).

E.g. LuaJIT 1.0.3 gets around 50% of the speed of fully-optimized and hand-tuned C code in the mandelbrot benchmark. That’s not so bad for a JIT engine that fits in only 25 Kilobytes IMHO.

]]>