6 ms·
1-4 corrected errors per month?! This sounds quite bad - which manufactor?
by xiconfjs 9y ago
1-4 corrected errors per month?! This sounds quite bad - which manufactor?
- kpil 9y agoIsn't that more or less in the range expected from background radiation bitflips when having 48 GiB?
- vel0city 9y agoI manage many servers with >=384GB RAM. They maybe detect an error once a year, with several never reporting an error in its life. If I noticed a server reporting errors more than twice a year I'd probably swap its memory, and if that didn't fix the problem I'd probably replace the server.
- StillBored 9y agoDepends on what you mean by "reporting an error". The default configuration for most machines I've used is to silently correct single bit failures unless they hit some threshold. So you actively have to run an EDAC (or similar) driver and poll the error counters. Its only with uncorrectable errors does it generally rise to the level of a full MCE.
- vel0city 9y agoMany server-grade motherboards will report ECC memory faults in the event log through IPMI.
- wolster 9y agoHow do these manifest themselves and how do you detect them?
- vel0city 9y agoThe error events are reported through IPMI in a system event log.
- ew6082 9y agoError rate would probably depend on the mass and quality of shielding above/around the servers.
- ChuckMcM 9y agoCrucial (six 8GB sticks). I don't put it into the 'bad' category. When overseeing 1600 machines in a data center with a collective 307TB of RAM suggests its actually on the 'good' side. Bad is multiple errors per day. Really bad is hundreds or thousands per day. Because they are corrected nothing actually happens on the system (other than what ever accessed them saw on the order of 700nS to access RAM rather than on the order of 100nS. If it gets a double bit error it will machine check so I'll know that my memory has failed me :-).
- sitkack 9y agoAll ram should be ECC (selective number of bits), then the processes that fab the ram can be made cheaper and sloppier. Zero bit-error-rate non-ECC RAM would actually be more expensive than ECC ram. A chip kill be on a much smaller granularity. I suspect these things haven't been done because the RAM cabal likes their high prices.