Re: Re-4: [ObjectWeb architecture] just an idea for (Java) native compilation

[Jean-Philippe Fassino <jeanphilippe.fassino@xxxxxxxxxxxxxxxxxxxx>]

Hi, look at my presentation of this morning to the Objectweb 
architecture meeting which
present some micro benchmarks between GCJ 3.3.1 and Sun JVM 1.4.1 (JIT 
& Interpreted).
(This benchmark which was run in the conditions you mention let 
thinking that HotSpot
is better in performance and don't consume more memory).

Best regards,

JP


Le vendredi, 26 sep 2003, à 09:57 Europe/Paris, <markus.karg@xxxxxxxxx> 
a écrit :

> Thank you for the links. Actually what that guy did was micro 
> benchmarking. This is unfair for java and other late-compiling VMs: 
> What that guy measures includes the startuptime of the VM, also 
> measures interpreted (!) code (not hot spot compiled code). To measure 
> hot spot compiled code, you need to put all functions into one, 
> execute that function one million times, then look the average value. 
> That's a real life test. No one cares on the java.exe startup time in 
> a server environment! No one cares the first 100 iterations of an 
> algorithm being run in iterpreted mode if the next 24 hours this 
> server will execute the same function one million time in hot spot 
> compiled mode.
>
> To write fair measuring micro benchmarks you need to know how hot spot 
> VMs work. That guy seems not to know (I told him today). So I'm 
> looking forward to see the actual results done with fair tests.
--
Jean-Philippe Fassino
France Télécom R&D/DTL/ASR
Ingénieur Recherche & Développement
28, chemin du vieux chêne
38243 Meylan Cedex
Tel : + 33 (0) 4 76 76 45 52
Fax : + 33 (0) 4 76 76 45 57
http;//www.francetelecom.com/rd

Attachment: 2003-09-ObjectWeb.pdf
Description: Adobe PDF document