[kaffe] Re: Re: Darwin/x86 134 of 144 tests Failed

Riccardo zuse at libero.it
Thu Sep 9 00:53:32 PDT 2004


In <413F74CA.7040608 at kaffe.org> Dalibor Topic  wrote:
> A very crude (and factually wrong, anyway, but as we are guessing 
> upper  bounds here, the advantages of RISC systems do not do harm) 
> estimation  would be to look at the lowest CPU megahertz counts people 
> are testing  on: if it takes 5 seconds with jit on a 200 MHz ix86 then 
> it may take  5*10*10 secs on a 20 MHz m68k with intrp.
>
> This would be of course error prone, and complex, and so on.

it would be an estimate, but quite wrong...

I have proven that my 50Mhz supsersparc had similar performances of the 
110MHz MicroSparcII
similar things are true for 68k and when comparing apples and oranges, 
say a 68k to a x86 things goes worse.
Also, while things seem more straightforward with the interpreter, JIT/
JIT3 react quite differently on different platforms and even OS specific 
stuff. [1]

So I think the heuristic would be too generic. Maybe something correct 
would be to run something in java (that would be feasible only after the 
build) and test with the currently built engine on that cpu some sort of 
KaffeBogoMips. While still very rough, it would have the advantage that 
it would take in account of the real speed of intrp/jit on that platform 
and not only of the CPU.

If that is needed I could adapt one of my benchmarks [1] to the purpose. 
So it could be run (supplied already in byte code) at the beginning of 
make check (supposing... kaffe is working enough to run it) and then 
write the estimate in some env variable. [2]

Also here were are speaking of a maximum time and as I wrote in another 
mail I would prefer a parameter to pass to make at check time; instead 
of complicate heuristics.

>From my experience with Guilhem, and from the description of your 
problem too Franz, I think most times timeouts problems are due to 
thread problems with that platforms, with threads locking up or dying or 
something similar.

The only sure thing I can ay is that on a MicroSparc II machine, with 
jthreads and JIT 300 seconds are not enough. But heck we know how slow 
our intrp is... If on the same box they fail with JIT, I'd suspect more 
technical problems.

-Riccardo

[1]: http://homepage.mac.com/riccardo_mottola/kaffe-devel/benchmarks/
index.html

[2]: I think that the auto-java check and the idea of passing 
"timeout=300" could be combined with some shell/make trickery. Maybe 
"timeout=auto" could run the KaffeBogoMips.





More information about the kaffe mailing list