Opteron

Petr Bezděk freebsd at ada-net.cz
Mon Oct 18 15:18:42 CEST 2004


Divacky Roman wrote:
> On Fri, Oct 15, 2004 at 03:33:27PM +0200, Jozef Babjak wrote:
> 
>>On Fri, Oct 15, 2004 at 03:26:39PM +0200, Petr Bezděk wrote:
>>
>>
>>>>># time make -j5 buildworld
>>>>>1662.749u 623.504s 20:16.41 187.9%      4326+3031k 3703+3899io 973pf+0w
>>
>>>>2) 20 minut na buildworld na 2xSMP je pomerne slabe... uz od vic lidi jsem
>>>>slysel ze na amd64 (iirc) maji pomale kompilace. mozna by bylo fajn to 
>>>>kapku vyprofilovat a podivat se co se deje

Nemam ted po ruce zelezo, abych mohl udelat objektivni srovnani, ale 
matne si vzpominam, ze kompilace 5.1-RELEASE trvala na P4 2.8 bez HT cca 
47 minut. Na opteronu s -j1 jsem dosahoval casu kolem 33 minut (jina 
verze gcc, asi i pribylo par radku kodu od 5.1-R). Asi je jeste co 
zlepsovat, ale oznacovat to jako "pomerne slabe" mi pripadne moc 
pritazeny za vlasy. Udelal jsem nejake mereni, prikladam je na konci 
mailu - omlouvam se za delku.

> mimochodem, ted mne napada -j5 ? ie. dva procesy na procesor + 1 ktery bude
> furt migrovat? zkus to (puvodni tazateli) s -j6 nebo -j4

-j5 mam nejak zazitou, kdyz kompiluji na Gentoo s distcc na dvou 
strojich = pocet procesoru x 2 + 1.

Jinak odhadnout pocet paralelnich spusteni make je trochu magie a 
vysledky meho testu me trochu prekvapily. Delka kompilace se vice mene 
nelisila, at jsem pustil buildworld s -j2 nebo s -j8. Jen load s vetsim 
cislem rostl a to je logicke.

Pripominka - kdyz spustim kompilaci se sudym poctem jobu, tak nikde neni 
napsano, ze prave polovina procesu pobezi na prvnim procesoru a druha 
pulka procesu na druhem procesoru. Takze procesy budou migrovat tak i 
tak z jednoho procesoru na druhy a jejich poradi a cas straveny na 
procesoru urcuje scheduler.

TEST 1:

kernel - SCHED_4BSD

make -j1 buildworld - Mon Oct 18 09:23:30 CEST 2004
real    33m19.685s
user    26m27.193s
sys     7m23.083s

make -j2 buildworld - Mon Oct 18 09:56:51 CEST 2004
real    19m5.724s
user    26m27.030s
sys     8m48.764s

make -j3 buildworld - Mon Oct 18 10:15:57 CEST 2004
real    19m17.120s
user    27m0.258s
sys     9m24.398s

make -j4 buildworld - Mon Oct 18 10:35:15 CEST 2004
real    19m28.596s
user    27m13.288s
sys     9m30.986s

make -j5 buildworld - Mon Oct 18 10:54:44 CEST 2004
real    19m31.183s
user    27m17.339s
sys     9m36.484s

make -j6 buildworld - Mon Oct 18 11:14:16 CEST 2004
real    19m32.553s
user    27m17.540s
sys     9m37.983s

make -j7 buildworld - Mon Oct 18 11:33:49 CEST 2004
real    19m34.345s
user    27m17.470s
sys     9m43.489s

make -j8 buildworld - Mon Oct 18 11:53:24 CEST 2004
real    19m40.486s
user    27m21.030s
sys     9m41.123s


TEST 2:

kernel - SCHED_ULE

make -j1 buildworld - Mon Oct 18 05:10:00 CEST 2004
real    32m29.167s
user    25m39.348s
sys     7m7.122s

make -j2 buildworld - Mon Oct 18 05:42:29 CEST 2004
real    19m41.893s
user    26m10.543s
sys     8m38.275s

make -j3 buildworld - Mon Oct 18 06:02:12 CEST 2004
real    19m37.473s
user    26m52.346s
sys     9m4.580s

make -j4 buildworld - Mon Oct 18 06:21:50 CEST 2004
real    20m1.596s
user    27m11.004s
sys     9m15.048s

make -j5 buildworld - Mon Oct 18 06:41:52 CEST 2004
real    19m49.487s
user    27m17.463s
sys     9m17.989s

make -j6 buildworld - Mon Oct 18 07:01:42 CEST 2004
real    20m11.110s
user    27m26.009s
sys     9m24.469s

make -j7 buildworld - Mon Oct 18 07:21:54 CEST 2004
real    20m13.680s
user    27m30.701s
sys     9m26.762s

make -j8 buildworld - Mon Oct 18 07:42:08 CEST 2004
real    20m8.097s
user    27m26.185s
sys     9m26.223s

-- 
Petr Bezděk



More information about the Users-l mailing list