On Oct 8, 2009, at 10:28 AM, Dan wrote:

> What got me to thinking is this article on ars...
>
> <http://arstechnica.com/business/news/2009/10/dram-study-turns-assumptions-about-errors-upside-down.ars
>  
> >



With the caveat that this is memory studied under high utilization  
regime; this is RAM in datacenter systems. The study showed that  
failure rate was proportional to utilization, and "About 8 percent of  
DIMMs were responsible for over 90 percent of the errors".

If only 8% of the systems display the error, is it worth the added  
costs of adding ECC RAM to the 90% of systems that don't have these  
failures? This is amplified by the findings that hard errors  
predominate. (which is the same thing as that 8% figure...these are  
faulty DIMMS, not cosmic rays)

Finally, the error rates only appear high because previously published  
error rates are so staggeringly (and artificially) low. In the entire  
study (and remember, these are 24/7/.999nnn uptime systems running at  
a high utilization all the time, hardly comparable to virtually any  
systems owned by people on this list) less than a third of the systems  
suffered *any* memory failures in a year.

SO to recap: Under datacenter utilization regimes in server systems  
(much more rigorous than the average computer user), 90% of all memory  
failures are caused by 8% of the DIMMS, and 66% of the systems saw no  
failures in an average year.

picking RAM out of a hat, Data Memory Systems sells 1G DDR2 PC2-5300  
667MHz DIMMS for the following prices:

Type            Price   %premium
Non-ECC         $23       --
Unreg ECC       $36      56%
Reg. ECC        $39      70%

The cost/benefit ratio just isn't there...the conclusions to be drawn  
from this article are that DIMM manufacturers need better testing and  
QC protocols to identify that 8%; charging 10% more per DIMM for  
adequate testing is much more likely to bring about real, widespread  
improvements than spending 60 or 70% more per DIMM for error correction.

Finally there's the issue of what the errors actually mean. What  
happens if one bit flips in a memory representation of a jpeg?

-- 
Bruce Johnson
University of Arizona
College of Pharmacy
Information Technology Group

Institutions do not have opinions, merely customs



--~--~---------~--~----~------------~-------~--~----~
You received this message because you are a member of G-Group, a group for 
those using G3, G4, and G5 desktop Macs - with a particular focus on Power Macs.
The list FAQ is at http://lowendmac.com/lists/g-list.shtml and our netiquette 
guide is at http://www.lowendmac.com/lists/netiquette.shtml
To post to this group, send email to [email protected]
For more options, visit this group at http://groups.google.com/group/g3-5-list
-~----------~----~----~----~------~----~------~--~---

Reply via email to