Re: Sarge, kernel-image, and i586

[Adam Majer <adamm@galacticasoftware.com>]

Andrew Suffield wrote:

>On Wed, Aug 25, 2004 at 12:23:54AM -0500, Adam Majer wrote:
>  
>
>>Matt Zimmerman wrote:
>>
>>    
>>
>>>On Mon, Aug 09, 2004 at 02:56:15PM -0500, Adam Majer wrote:
>>>
>>> 
>>>
>>>      
>>>
>>>>There is no more 586 kernel though. Not even the regular one. The 386
>>>>version is quite slower on a Pentium MMX than an optimized one.
>>>>   
>>>>
>>>>        
>>>>
>>>Have you measured this?
>>>
>>> 
>>>
>>>      
>>>
>>With my stop watch. Really. KDE was at least twice as slow with a 386
>>kernel than a 586mmx kernel. There is a *huge* performance penalty if
>>one is running a 386 kernel. I have tried this on a Pentium 200 MMX with
>>196MB. With slow machines like this, the difference between a 386 and
>>586 can be seconds to start things like terminal or Konqueror. It is
>>actually more than a minute booting up. The system is essentially
>>unusable with a 386 kernel, but it is still running at a steady pace
>>under a 586mmx kernel.
>>    
>>
>
>This is a *5*86 chip, and not a 686, right? This is not implausible;
>the 586 was a *disaster* in hindsight, and an oddball in an otherwise
>tidy series of chips. I believe the problem lies with the pipeline
>scheduling; at any rate, you want 586-targetted code on a 586 and
>486-targetted on anything else. You also want to throw your 586 in the
>bin.
>  
>
Yes, a 586 (Pentium MMX). It is still a usable desktop box for basic
things like word processing or email.

>On the other hand, there are at least half a dozen other things that
>could possibly be responsible for this, and it doesn't sound like
>you've ruled any of them out. Certainly this is not compelling
>evidence; you'd have to analyse exactly *why* it made a difference.
>  
>
Well, 386 kernel makes the box run slow. 586mmx kernel (compiled from
Debian kernel sources - changed 386 to 586mmx in config, no other
changes) runs twice as fast.

There were no other changes.


>>I'm not complaining that there is no 2.6.x 586 or 586-mmx kernels since
>>I can just rebuild one. But there is a huge performance loss running a
>>386 kernel on anything else than a 386 or maybe a 486 (but I haven't
>>tried a 486).
>>    
>>
>
>This, however, is *not* true. If you have a *6*86 then the design flaw
>was fixed and the problem does not occur. What's more, a 586 isn't
>much happier about 686-targetted code than it is about 386-targetted
>code. And nothing else likes running 586-targetted code - it can be
>slower than 486-targetted code in some circumstances, when running on
>a 686. The difference between 486-targetted and 686-targetted, on a
>686, is small to nonexistent.
>  
>

I know that 586 targeted code sucks on 686 (PentiumPro) and newer chips.
The problem is that 586 was designed to run 16-bit, not 32-bit code.

- Adam

-- 
Building your applications one byte at a time
http://www.galacticasoftware.com