[Beowulf] Not quite Walmart, or, living without ECC?

[David Mathog]

Joe Landman wrote:

> Memtest and fellow travelers access memory in a very regular manner. 
> Which is unlike the way most programs access memory.

Hmm, I see what you mean (although in my code I can often go through
memory in exactly the same "regular" manner).  If the memory tester
doesn't specifically disable CPU cache (to force direct write through
and read through) then it's going to be moving lines of memory at a
time, which actually doesn't load the non-cache memory very much.  For
the purposes of memory testing, beyond just looking for an overtly bad
memory location, it would probably make more sense to go "the other
way" through memory, forcing as many cache misses as possible.  That
would result in fewer passes per unit time for the testing software, but
each test is more likely to cause a problem, since the memory is busier.  

I'll ask the memtest86+ folks which way they go about it.  If they are
going straight through memory that might explain some of the differences
that have been reported here.

Regards,

David Mathog
mathog at caltech.edu
Manager, Sequence Analysis Facility, Biology Division, Caltech