Hacker News Reader

Comment by jacquesm | Hacker News Reader

The frequency of these errors should be enough to cause you to distrust any kind of information output by a computer without ECC.

Edit: thinking about this a bit longer: that frequency is actually so high that you may well have a broken module in there. Note how it is the same module and the same address every time.

View on HN

Seems like a semi-stuck bit, it'd definitely cause issues w/o ECC but seems to chug along with the circuits doing their job. Best would probably be to add an memory-range exclusion to the kernel at boot to avoid that single area since the sticks seems good otherwise.

I completely forgot you could do that.

Ime ram is either bad or good. I’ve had ecc errors like this and I always ask the DC to replace the ram. After that, 0 errors forever. Same reason why I’m confident a 24 hour memcheck is sufficient for non ECC ram.

Generqlly the same here, but I have had sticks fail after some time in use. I had to rma the ram in my frame.work laptop after it failed. No reason or clue why but it happened after 6 minths or so. No issues with the rma though and it went fine with no issues since. ecc if it was supported there might have given me a heads up about it and avoided needing me to restore from backups when the fs corrupted.

I am fully aware the module is not 100% working, i.e., it is faulty at a specific physical address. That's OK for my personal desktop though, unless the condition worsens, and UCEs (which will panic my kernel) follow.

Any chance it's something that's being affected by temperature?

Along the lines of "computer gets toasty doing work, ecc errors start happening"?

Stuff like that could just mean the memory sticks need pushing in a bit more.

I tried to test and control for that to the best of my ability, but ambient/operating temperature does not seem to be part of the equation.

It'd be cute to log what processes are using that memory at time of error. Fun to speculate about whether a kernel bit flip is better or worse than ones in a web browser, photo editor, spreadsheet, network storage client...

You could test that empirically by setting up a box with the express intent to crash it and then using a chaos monkey like mechanism where you start injecting single bit faults into memory at random addresses. Wonder how long the box would be up before you start noticing something is broken. It would be funny if you accidentally killed the chaos monkey first! Best not use that box for banking...

Reply on news.ycombinator.com