Re: pointer questions

["Dr. Adrian Wrigley" <amtw@linuxchip.demon.co.uk.uk.uk>]

On Mon, 31 Oct 2005 06:21:37 +0000, Jeffrey R. Carter wrote:

> Robert A Duff wrote:
> 
>> That's hard to believe.  Dr. Wrigley said the hardware failures turned
>> from "insideous" to "catastophic" when he changed some sort of Things to
>> pointers-to-Things.  I take that to mean, he got wrong answers before,
>> and crashes after.  Is that right, Dr. Wrigley?

I Thinks so, but the error rate was very low, so it is hard to tell.

> Thinking about it a little more, I get the idea that before he had 1 pointer to 
> a 1 GB data structure on the heap. Then he changed to a large number of pointers 
> to smaller data structures on the heap, equivalent to the previous 1 GB structure.

This is exactly the situation.  The "Things" were about 36 bytes each
and I changed to having around 20 million pointers to things.
Single bit errors in pointers had a much more significant effect
than single bit errors in "Things" (which tended to be ignored
for various reasons).

> With more pointers, there's a greater chance that one of the pointers might have 
> a bit flip, resulting in the occasional crash (1 crash every few months, IIRC).

Yes.

> Of course, this is sheer speculation.

well speculated! (was it so unclear?)

Another "feature" I observed was that files could stay cached by the OS
for months, and accumulate the occasional single-bit error.  But when you
evict the cached pages and read the data again, the errors disappear.
Plenty of scope for very rare Heisenbugs.

The warning I give is that these error rates are "normal" for modern SDRAM,
but aren't usually noticed because they usually only show up if
you have several GB of memory and care about every bit 24/7.
Operating a financial business, I care about this!
-- 
Adrian