| View previous topic :: View next topic |
| Author |
Message |
Cyrack
n00b
Joined: 15 Jun 2003
Posts: 39
Location: Denmark (Herning)
|
 Posted: Thu Jan 27, 2005 11:30 pm Post subject: Acovea/GCC |
 |
|
Hi, just stumbled over something interesting (or mayby I just making a fool of myself right now). Anyway: I just tried to run acovea on my system, first I tried with my own config (just altered the opteron config so the -march would read athlon64 instead of opteron), the final output was:
| Code: |
optimistic options:
-fno-guess-branch-probability (1.378)
-fno-if-conversion (1.063)
-fschedule-insns (2.325)
-frename-registers (1.288)
-fprefetch-loop-arrays (1.288)
-fpeel-loops (1.559)
-mieee-fp (2.009)
-mfpmath=387 (1.333)
-funsafe-math-optimizations (2.37)
-fno-trapping-math (1.153)
-ffinite-math-only (1.198)
pessimistic options:
-fno-crossjumping (-1.146)
-fschedule-insns2 (-1.371)
-fregmove (-1.957)
-falign-labels (-1.101)
-ffloat-store (-1.822)
-fnew-ra (-1.642)
-funroll-loops (-1.552)
-funroll-all-loops (-1.687)
-fbranch-target-load-optimize2 (-1.281)
-mfpmath=sse (-2.093)
-mfpmath=sse,387 (-1.371)
-fno-math-errno (-1.011)
|
Now I tried using the original opteron config and I got this output:
| Code: |
optimistic options:
-fno-crossjumping (1.365)
-fschedule-insns (2.513)
-freorder-functions (1.25)
-funit-at-a-time (1.308)
-fweb (1.021)
-fprefetch-loop-arrays (1.365)
-mieee-fp (1.939)
-funsafe-math-optimizations (3.029)
pessimistic options:
-fforce-mem (-1.103)
-fregmove (-1.619)
-ffloat-store (-1.332)
-fmove-all-movables (-1.103)
-fnew-ra (-1.791)
-funroll-all-loops (-1.504)
-fbranch-target-load-optimize2 (-2.078)
-mfpmath=sse (-2.595)
-fno-math-errno (-1.275)
-fno-trapping-math (-1.045)
|
The tests was done during night time, with all cron-jobs disabled, no network connected and noone using the computer during the test, to avoid any load-differences. The benchmark was done with almabench.c. Before the start of the tests I restarted the computer, and wiped out any tmp-dirs so the entire disk would seem unchanged.
Now, here is the tricky part: why does the output differ so much? According to the gcc manual the athlon64 and opteron flag is identical, and therefor should have absolutly no difference on the test result!
(The only thing which could be different is the fact I have Cool and Quiet activated, and therefor the clock could have been different, but then again, the two tests should have put the same amount stress on the system, causing no difference) |
|
| Back to top |
|
 |
moocha
Watchman
Joined: 21 Oct 2003
Posts: 5722
|
| Posted: Thu Jan 27, 2005 11:51 pm Post subject: |
|
|
Ah, the wonders of non-deterministic computing.
I've personally encountered trouble that finally got traced to higher than normal air humidity in the machine room at night, so I shouldn't wonder too much at these results.
_________________
Military Commissions Act of 2006: http://tinyurl.com/jrcto
"Those who would give up essential liberty to purchase a little temporary safety deserve neither liberty nor safety."
-- attributed to Benjamin Franklin |
|
| Back to top |
|
|
Cyrack
n00b
Joined: 15 Jun 2003
Posts: 39
Location: Denmark (Herning)
|
| Posted: Fri Jan 28, 2005 1:16 am Post subject: |
|
|
| Well, now that is something to optimize against.... Change your CHOST-flags when room-temperature drops by 3 degrees or when humitity increases by 5%.... Could cause a hell lot of compiling.... |
|
| Back to top |
|
|
curtis119
Bodhisattva
Joined: 10 Mar 2003
Posts: 2160
Location: Toledo, Ohio,USA, North America, Earth, SOL System, Milky Way, The Universe, The Cosmos, and Beyond.
|
| Posted: Fri Jan 28, 2005 2:04 am Post subject: Re: Acovea/GCC |
|
|
| Cyrack wrote: | | Hi, just stumbled over something interesting (or mayby I just making a fool of myself right now). |
I vote for fool. But that's just me. 
_________________
Gentoo: it's like wiping your ass with silk. |
|
| Back to top |
|
|
FastTurtle
Guru
Joined: 03 Sep 2002
Posts: 477
Location: Flakey Shake & Bake Caliornia, USA
|
| Posted: Wed Jun 08, 2005 3:46 am Post subject: |
|
|
What caused the difference should be obvious: Random choice of starting flags. There is no way to get absolutely the same flags every time due to the random nature of the start.
IMO: The only way to use acovea effectively is to run it multiple times (10x min) and examine the output to determine what flags have a better change of showing up. For example, if a flag shows up across several runs (6x min) I would assume that it's one that should be included system wide as it has a higher probability of improving things on that system. I would most definately include any flag that shows up in (8x) runs due to probability.
I am actually planning on doing such a multiple run test on my fresh stage1 install after I have determined my initial cflags as I'm currently running acovea.
_________________
AsRock B550 Phantom Gaming 4
128GB 3200 Mhz memory
4x 4TB Sata - 2x 2TB Sata SSD - 4x 450GB SaS - 3x 900GB SaS - 72GB SaS for Boot
LSI 9211-8i in HBA mode for all of the SaS drives
Radeon 6800 (Non XT) for GPU |
|
| Back to top |
|
|
|
|
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
|
|
Gentoo on AMD64
