[PATCH 0/4] 8xx: Optimize TLB Miss code.

Joakim Tjernlund joakim.tjernlund at transmode.se
Fri Mar 5 21:40:29 EST 2010


Heiko Schocher <hs at denx.de> wrote on 2010/03/04 17:30:07:
>
> Hello Joakim,
>
> Joakim Tjernlund wrote:
> > Wolfgang Denk <wd at denx.de> wrote on 2010/03/04 13:16:56:
> >> From: Wolfgang Denk <wd at denx.de>
> >> To: hs at denx.de
> >> Cc: Joakim Tjernlund <joakim.tjernlund at transmode.se>, Klaus-Jürgen
> >> <heydeck at kieback-peter.de>, linuxppc-dev at ozlabs.org, Scott Wood
> >> <scottwood at freescale.com>
> >> Date: 2010/03/04 13:17
> >> Subject: Re: [PATCH 0/4] 8xx: Optimize TLB Miss code.
> >>
> >> Dear Heiko,
> >>
> >> thanks for running the tests.
> >>
> >> In message <4B8F8BB4.6070201 at denx.de> you wrote:
> >>> here the results:
> >>>
> >>> run   version
> >>>
> >>> 1-4   2.6.33-rc6 without your patches
> >>> 5-8   2.6.33-rc6 with all your patches
> >>> 9-12   2.6.33-rc6 with patches 1,2 and 4 (without 8xx: Don't touch ACCESSED
> >> when no SWAP)
> >>> 13-16   2.6.33-rc6 with all your patches and CONFIG_PIN_TLB=y
> >> So CONFIG_PIN_TLB imroves the performance as expected, while the other
> >> patches don;t show any measurable improvememt - or am I reading the
> >> results incorrectly?
> >
> > Close but not quite. What stands out most is:
> >
> > Memory latencies in nanoseconds - smaller is better
> >     (WARNING - may not be correct, check graphs)
> > ------------------------------------------------------------------------------
> > Host                 OS   Mhz   L1 $   L2 $    Main mem    Rand mem    Guesses
> > --------- -------------   ---   ----   ----    --------    --------    -------
> > tqm8xx    Linux 2.6.33-    66   31.8  141.0       184.0      1165.7
> > tqm8xx    Linux 2.6.33-    66   31.8  141.2       184.2      1165.3
> > tqm8xx    Linux 2.6.33-    66   31.8  141.3       184.3      1165.6
> > tqm8xx    Linux 2.6.33-    66   31.8  141.3       184.2      1166.2
> >
> > tqm8xx    Linux 2.6.33-    66   31.8  141.0       171.8      1100.5    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  141.0       171.8      1102.5    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  141.0       171.8      1101.7    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  141.0       171.8      1101.6    No L2 cache?
> >
> > tqm8xx    Linux 2.6.33-    66   31.8  141.1       173.4      1149.1    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  141.1       173.4      1149.0    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.7  141.1       173.4      1148.7    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.7  141.1       173.4      1148.2    No L2 cache?
> >
> > tqm8xx    Linux 2.6.33-    66   31.8  171.1       171.7      1099.8    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  171.1       171.6      1100.5    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.7  171.0       171.7      1101.0    No L2 cache?
> > tqm8xx    Linux 2.6.33-    66   31.8  171.0       171.6      1101.3    No L2 cache?
> >
> >
> > Besides the numbers, note how the first group doesn't have a Guesses entry.
> > Is there something odd with the results for the first group?
>
> Hmm.. just to be safe, I made this test again, but it shows also no entry in
> "Guesses" ... Hardware, Linux Source, rootFS, lmbench sources, all the
> same ...

OK

>
> > Also, since you are using MODULES, patch 2 is nullified.
> > Patch 1 is very minor and should not show I think.
> > This leaves patches 3 & 4.
> > There appears to be something funny with patch 3,Don't touch ACCESSED when no SWAP, as
> > it yields bad numbers for Prot Fault so perhaps I am missing something that
> needs ACCESSED
> > even if NO_SWAP. Perhaps a someone that knows MM in Linux knows?
> > Is there any messages in the kernel log(dmesg)?
>
> I couldn;t find something in the output with dmesg ... but if you
> want this output, I can send it to you.

No, if you can't find anything in there, I won't either.

What would be interesting is to skip patch 3 and turn off
MODULES add PIN_TLB and compare that against your unpatched .33 but
with MODULES off and PIN_TLB on

 Jocke



More information about the Linuxppc-dev mailing list