From: Godmar Back (gback@cs.utah.edu)
Date: Tue Jan 05 1999 - 11:15:01 EST
FYI:
Forwarded message:
> From gback@cs.utah.edu Mon Jan 4 23:59:28 1999
> From: Godmar Back <gback@cs.utah.edu>
> Message-Id: <199901050700.AAA29111@lal.cs.utah.edu>
> Subject: Re: Curious Kaffe vs. jdk speed test results under Linux (fwd)
> To: dej@inode.org
> Date: Tue, 5 Jan 1999 00:00:10 -0700 (MST)
> X-Mailer: ELM [version 2.4 PL25]
> MIME-Version: 1.0
> Content-Type: text/plain; charset=US-ASCII
> Content-Transfer-Encoding: 7bit
>
>
> Btw, the time it takes to jit the 174 methods in this test is negligible.
> The reason kaffe sucks so much is because it spends more time in gc and
> because the quality of its jitted code is so low.
> Plus, I just fixed another bug where kaffe would invoke its gc twice in
> a row for no good reason other than two threads (main & finalizer) asked
> for it. The numbers below already reflect that fix.
>
> I was able to get it under jdk's time, though, when I allowed for 64MB
> which didn't trigger any gc.
> Basically:
>
> Sun with no gc:
> 6545 ms
> Sun out of the box (does about 20 small gcs)
> 6913 ms (this is actually very good!)
>
> Kaffe with no gc (with -ms 64M):
> 5423 ms
> FACILITY USER TIME SYSTEM TIME CALLS
> --------------- --------------- --------------- ---------------
> JIT 0.020000 0.000000 178
> TOTAL 5.360000 0.280000
>
> Kaffe with 1 gc: (with -ms 32M)
> 7223 ms
> FACILITY USER TIME SYSTEM TIME CALLS
> --------------- --------------- --------------- ---------------
> gc-sweep 0.330000 0.000000 1
> gc-scan 0.980000 0.000000 1
> JIT 0.070000 0.000000 180
> TOTAL 7.180000 0.280000
>
> Kaffe out of the box:
> 8650 ms
> FACILITY USER TIME SYSTEM TIME CALLS
> --------------- --------------- --------------- ---------------
> gc-sweep 0.480000 0.000000 41
> gc-scan 2.460000 0.000000 41
> JIT 0.040000 0.000000 180
> TOTAL 8.820000 0.060000
>
> We need a better gc and a better strategy to invoke it.
>
> - Godmar
>
This archive was generated by hypermail 2b29 : Sat Sep 23 2000 - 19:57:29 EDT