non-ECC or ECC RAM?

[Blind_Sniper]

Hi!
I thinking of building new box and can't decide which RAM to buy: ECC or non-ECC?
Somewhere on web I read that non-ECC RAM can cause segfoults when compiling huge projects if using more than 16 Gb of RAM (32 or 64Gb). And there was recommendations to buy ECC RAM if you need 32Gb or more.
Anyone here using non-ECC RAM of 32 Gb or higher?

[mike155]

You may want to read: https://www.phoronix.com/scan.php?page= ... rvalds-ECC

[NeddySeagoon]

Blind_Sniper,

ECC costs you one clock cycle to do the check and correct if required.
If you do get a detected error it will be logged in dmesg.

If no errors are corrected, you can be sure you didn't need ECC RAM. Without it, you never know.
ECC will detect and correct single bit errors, detect but not correct two bit errors and after that all bets are off.

Do you feel lucky?

[Ant P.]

It's a constant gamble with physics.

If you don't want the cost of ECC, the next best thing you can do is overprovision your RAM so it's never highly occupied with important data. It *will* go wrong at some point, and the best you can hope for then is that the bit getting flipped is in unused space.

I'm doing the latter. 64GB non-ECC, only usually filled up when I'm running emerge.

[Blind_Sniper]

Blind_Sniper,

ECC costs you one clock cycle to do the check and correct if required.

I know it will be slower than non-ECC for about 10%. The thing is availability of high speed ECC RAM. I want to buy 2x16Gb 3600 Mhz or 4000 Mhz, but there is no such a fast ECC RAM. At least I didn't see any of it.

If you don't want the cost of ECC, the next best thing you can do is overprovision your RAM so it's never highly occupied with important data. It *will* go wrong at some point, and the best you can hope for then is that the bit getting flipped is in unused space.

I'm doing the latter. 64GB non-ECC, only usually filled up when I'm running emerge.

So, it works for you?

@mike155, thanks for the link, but I think, my use case differs of use case Linus cares about (high load servers)

[NeddySeagoon]

Blind_Sniper,

I have access to two arm64 systems, both with 128GB RAM. One reports ECC errors several time a month.
Its in a data centre.
The other has yet to report any errors.

If you want RAM at 3600 Mhz or 4000 Mhz, why would you increase the latency with ECC?

I guess you want unregistered DDR4 too?
It boils down to

You cannae break the laws of physics captain

[Blind_Sniper]

Several errors for 128Gb when it runs for a month?
If so my use case (desktop, browsing, playing chess online and weekly running emerge) with max uptime couple of days doesn't require any ECC.
I started this topic because I read some complain that running compiler at 32Gb of non-ECC RAM caused segfolts in 20 mins. May be that dude had a different problem? Though he stated that problem was fixed bby replacing non-ECC with ECC modules.

So I guess, I will pick non-ECC (unregistered, afaik only ECC RAM can be registered).

[NeddySeagoon]

Blind_Sniper,

... and none at all after several months on the other system.

[Tony0945]

I started this topic because I read some complain that running compiler at 32Gb of non-ECC RAM caused segfolts in 20 mins. May be that dude had a different problem? Though he stated that problem was fixed bby replacing non-ECC with ECC modules.

Maybe he just had bad RAM or was overclocking it unmercifully. I buy RAM from high quality manufacturers and don't push the specs. I had one stick of Ballistix RAM fail after five years. Crucial honored the lifetime warranty but no longer made that Ballistix RAM. They offered a credit against direct purchase or my taking twice as much standard RAM (also with lifetime warranty) . I took lthe latter offer. Of course, RAM doesn't last forever. They probably have calculated that most buyers will junk their machine for a newer model before the RAM fails.

[NeddySeagoon]

Blind_Sniper,

ECC is a safety net. One random soft error every now and again is probably a cosmic ray.
If you need the ECC on a regular basis, then something is wrong and needs to be fixed.