3 ms·
> So the probability of having at least one bit error in 4 gigabytes of memory at sea level on planet Earth in 72 hours is over 95% . This is misleading. A fla
by barrkel 17y ago
> So the probability of having at least one bit error in 4 gigabytes of memory at sea level on planet Earth in 72 hours is over 95% .
This is misleading. A flaky machine will indeed see bit errors, and it will probably be visible as random crashes, but that's not even necessarily the case for the average machine. If you look at the quantitative study from Google, which the author links to:
http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf
... you can see that in terms of errors per DIMM:
"Across the entire fleet, 8.2% of all DIMMs are affected by correctable errors and an average DIMM experiences nearly 4000 correctable errors per year. These numbers vary greatly by platform. Around 20% of DIMMs in Platform A and B are affected by correctable errors per year, compared to less than 4% of DIMMs in Platform C and D. Only 0.05-0.08% of the DIMMs in Platform A and Platform E see an uncorrectable error per year compared to nearly 0.3% of the DIMMs in Platform C and Platform D. The mean number of correctable errors per DIMM are more comparable, ranging from 3351-4530 correctable errors per year."
So the mean rate of correctable errors is high, but the variance is also very high: depending on the manufacturer, 80 to 96% of DIMMs see out a whole year without a single correctable error. If the original statistics of 95% chance of error in 3 days were correct, a single-DIMM machine ought to have approximately 0% (astronomically close to 0%) of living out a whole year without errors - but we can see here that you seem to have between 80 and 96% of single-DIMM machines doing just that.
The moral here is to test your memory for a while - preferably a few days - before trusting the DIMMs. But once you know you have good DIMMs, it doesn't look like you need to be quite so paranoid about bit errors.
- scotty79 17y agoI don't think that there was any mention that error rate was dependent on manufacturer.
- tbrownaw 17y ago> But once you know you have good DIMMs, it doesn't look like you need to be quite so paranoid about bit errors. Assuming that only the one-error-per-year cases were due to random bit flips, and all the multiple-errors-per-year cases were due to bad DIMMs, I came up with about a 1/5 chance of getting a single random bit-flip over a 6 year lifespan. But there also seems to be about a 1/3 chance of having a DIMM randomly go bad after a couple years, which of course without ECC would manifest as random crashes and lost (or maybe corrupted) work.
- Estragon 17y agoSeems like running memtest every six months or so would be a good policy. It would be nice if there were a way to test the memory while a machine is running.
- regularfry 17y agoI remember someone-somewhere saying that in their experience building GCC was a more thorough memory test than memtest itself...
- psranga 17y agoThe errors we're talking about here are transient. The memory location itself is still usable, the contents get changed when a cosmic ray hits. After the hit, the corrupted value is held without a problem. Memtest checks if the memory location has a gross fault which prevents it from storing values correctly. Doesn't seem like memtest will help.
- Estragon 17y agoGood point. Thanks.
- anamax 17y ago> But once you know you have good DIMMs, it doesn't look like you need to be quite so paranoid about bit errors. During the talk, http://www.stanford.edu/class/ee380/fall-schedule-20092010.html http://www.stanford.edu/class/ee380/fall-schedule-20092010.h... , IIRC, she said that good DIMMs went bad over a fairly short period of time, on the order of 2-3 years.