John R Pierce wrote:
I do think a 64bit version of memtest86 would be a good thing... I'm unclear how a 32bit app could test memory above about 3GB unless its playing the PAE game (3GB because in a 32bit system, the PCI bus peripherals take about the top ~ 1GB for their I/O space, so the ram has to be mapped to a higher physical address, only accessible via PAE...
Yea, addressing is one issue, but I have seen problems on 1GB machines too. Mind you, memtest is probably right and the memory modules probably were ok. It is some combination of chipsets, memory and linux in 64-bit mode that stress parts of the entire memory system. That is what memtest is unable to detect. Few people run in 64-bit mode compared to 32-bit... still some snags to work in the chipsets/memory I guess.
In real life a "while true; compile kernel; done" loop would panic the kernel in 10-30 minutes. Memtest didn't find problems. Changing memory modules to a different brand or lowering clockspeed made the system stable.
As for 32-bit and testing... remember there is a big difference between address space and memory. You would just map the memory in the 4GB address range and test it, same method as you would do with PAE mapping - up to 64GB. But that is still big enough for most people :)