Hacker News Reader

Comment by Aeglaecia | Hacker News Reader

im sure that any reasonably charismatic software engineer could scare the shit out of a judge/jury based on code analysis ... perusing the linked nasa document, recurrence odds of this particular failure mode do seem significantly greater than the odds of a cosmic bit flip ...

View on HN

> Studies by IBM in the 1990s suggest that computers typically experience about one cosmic-ray-induced error per 256 megabytes of RAM per month.

https://www.scientificamerican.com/article/solar-storms-fast...

Just to give perspective on the bit flip probability. ECC ftw!

Is this was true, wouldn’t any modern computer would be crashing several times a month? And please don’t tell me “oh but it is”, because it is not.

How many of those errors vould result in a full system crash, though? And how many of them are just going to cause silent and mostly-harmless data corruption?

After all, was the error in the first line a typo on my side, or a single-bit upset?

A while ago some researchers registered off-by-one-bit domain name typos, which due to physical key positioning were unlikely to be the result of genuine mistyping. I can't find a reference right now, but I recall them getting quite a lot of queries!

Most of the RAM may not be critical enough to crash the whole system. Just some random app you have open or a browser tab. So even if it is true, most bit flips should not crash a system.

Yes, I know that. So why aren’t my applications or tabs crashing at least once a day?

No clue, but you can most likely simulate this for a process on linux. Actually, I might just try that and see what happens.

Write a Bit Flip simulator and report your observations ;)

Servers with ECC generally report zero recoverable memory errors until the chip starts failing, at which point there are increasingly many. Therefore the average server experiences zero cosmic ray related memory errors during its lifetime, despite having many times more memory than 256MB.

It's not true.

I have left memtest86+ running on a few dozen GB of memory for several days during burn-in testing, definitely more than enough to pass the "once per 256MB per month" threshold, and did not encounter any errors.

Reply on news.ycombinator.com