Volker Armin Hemmann <[email protected]> posted
[email protected], excerpted below, on  Thu,
19 Feb 2009 15:46:56 +0100:

> Get a known good psu and let memtest run for a couple of hours. If it
> throws errors, try to increase the voltage of the memory modules a tiny
> little bit.

Also note that memtest doesn't always catch errors, depending on what/
where they are in the memory subsystem.  I had a couple sticks that 
memtested good, that were borderline on their rated timings.  At the 
rated PC3200 speed they would sometimes return incorrect data.  When Tyan 
came out with a BIOS that would allow memory underclocking (the original 
didn't), I was able to underclock it to PC3000 speeds, and at that, even 
decrease a lot of the wait-states from standard so the speed was much 
closer to the original claimed 3200, and it was 100% solid.  But at 3200 
it simply wasn't stable.

But memtest doesn't stress timings, only the memory cells themselves, and 
they were fine, so memtest came back 100% fine as well.

FWIW, as long as I wasn't stressing the system, it would often stay up 
for days.  I could often complete "short" emerges, but had to devise a 
way to let emerge continue working on longer ones after a system crash, 
because it just couldn't get thru say a gcc or glibc full merge 
(effectively two builds, 32-bit and 64-bit, since I was still multilib at 
the time), or kdelibs or any of the monolithic kde modules, for that 
matter.  So I had a terrible time getting the system up and running the 
first time!

It would also fail bunzip2s fairly frequently, with checksum errors.

But memtest gave it 100% pass!

Eventually I upgraded memory, to 8 gigs worth (of registered DDR-1, not 
cheap!)!  The new memory worked fine without underclocking.

The moral is, memtest may catch memory cell errors, but it won't always 
catch memory timing errors, so if it says the memory is bad, replace it, 
but just because it says it's fine doesn't mean it actually is.  It 
doesn't have many false-positives (fails), but it can have false-
negatives (passes).

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman


Reply via email to