[linux-elitists] Dual bore - a question on diagnosis

[Phil Mayers]

On Thu, Dec 21, 2006 at 02:21:06PM +0000, Phil Mayers wrote:

>>Also, if the system has at least a gigabyte of RAM installed, try
>>passing:
>>
>>mem=1000m
>>
>>or some lower value (keep it less than 1 gig) to the kernel at the GRUB
>>or LILO command line.  I have seen cases (with first generation Centrino
>>laptops) where this would eliminate extreme sluggishness in computers
>>with >= 1 gig of RAM and a CONFIG_HIGHMEM=y kernel.
>
>Both excellent suggestions, thanks.
>
>I'm not in the office today so haven't tried them myself, but a 
>colleague tells me they're not effective. Grumble. See later post about 
>comedy reasons I'm not able to return it...

Hmm.

A combination of arguments seems to help things along:

mem=512m notsc ide0=noprobe ide1=noprobe

The MTRRs on the machine look wrong to me - the top 32meg is left 
uncached, but fixing them manually didn't seem to boost performance any.  
That could be because IDE/SATA problems were overshadowing it.

I've also discovered the FC6-64 kernel won't come up without the Nvidia 
card in - the onboard intel i965 graphics evidently is not friendly.

Dave suggested "cpufreq.debug=7" which seems to be behaving during the 
install at least - not surprising because the governor is set to 
"userspace" and left unhooked, so it'll just run at full tilt.

It has been suggested to me that windows XP has sporadic problems on the 
same hardware, and that MS KB 896256 references a combination of:

 * tsc synch (or rather, not), damaging the performance counters and 
thus making the scheduler do the wrong thing

 * acpi c-state promotion and demotion issues

 * processor power saving being enabled due to confusion over 
whole-package versus core-specific utilisation

...is to blame.

Thanks for all the awesome suggestions. I'm off for a couple of weeks 
now, if I learn anything useful I'll let you know.

Cheers & a happy holiday,
Phil