Samsung 980 NVMe SSD Linux Performance

[nils_]

I just replaced two 980 PRO with SN850ies. I noticed the performance is a lot better when re-formatting with 4KiB sector size using the NVME CLI.

[tildearrow]

I should point out that Bad Things Happen when Linux encounters full storage devices. That's why it still isn't a good idea to put all your data on a single partition, regardless the defaults on the major distributions. Bugs and misconfiguration happen. This has become even more the case with SSDs which require enough free cells to (hopefully) move data from failing cells to unused ones. Some MFGs are good at over provisioning. Some are not. I don't think it's wise with the obvious new race to the bottom to test out with your important data just who is going to win that race. It's also one reason spinning rust isn't going away any time soon. Static long lived data is better off on a traditional hard drive as they're generally better with refreshing the surface than SSDs which may or may not bother to check if the charge in the storage cells need a refresh. Testing is showing SSD cell charge may not be as stable as MFGs claim.

Why couldn't SSDs move used cells around when storage is full, instead of writing to the very few free cells?

[Royi]

With ZoL, the regular scrub of the pool does precisely this, since every block is checksum'd on ZFS.

So, Could you share more about your backup routine?
I have files I want to be sure are well backup (Family Photos / Videos).
I'm concerned which mechanism could validate them over time (That the HD itself didn't miss a bit flip etc...).

[onlyLinuxLuvUBack]

I should point out that Bad Things Happen when Linux encounters full storage devices. That's why it still isn't a good idea to put all your data on a single partition, regardless the defaults on the major distributions. Bugs and misconfiguration happen. This has become even more the case with SSDs which require enough free cells to (hopefully) move data from failing cells to unused ones. Some MFGs are good at over provisioning. Some are not. I don't think it's wise with the obvious new race to the bottom to test out with your important data just who is going to win that race. It's also one reason spinning rust isn't going away any time soon. Static long lived data is better off on a traditional hard drive as they're generally better with refreshing the surface than SSDs which may or may not bother to check if the charge in the storage cells need a refresh. Testing is showing SSD cell charge may not be as stable as MFGs claim.

Have you tried to recover data loss on a hdd ?
Its better on tape, bluerays, and encrypted cloud zip and hdd. Just make many copies instead of rely on hdd which will be 1000 dollars to recover.

[stormcrow]

I should point out that Bad Things Happen when Linux encounters full storage devices. That's why it still isn't a good idea to put all your data on a single partition, regardless the defaults on the major distributions. Bugs and misconfiguration happen. This has become even more the case with SSDs which require enough free cells to (hopefully) move data from failing cells to unused ones. Some MFGs are good at over provisioning. Some are not. I don't think it's wise with the obvious new race to the bottom to test out with your important data just who is going to win that race. It's also one reason spinning rust isn't going away any time soon. Static long lived data is better off on a traditional hard drive as they're generally better with refreshing the surface than SSDs which may or may not bother to check if the charge in the storage cells need a refresh. Testing is showing SSD cell charge may not be as stable as MFGs claim.

[torsionbar28]

The question is what checks regularly that no bit was flipped in data?

With ZoL, the regular scrub of the pool does precisely this, since every block is checksum'd on ZFS.

[Royi]

I made the switch only last year. The key is to use mdadm (or ZoL) and create a mirror. Plus regular backups to some external media. Downside of course, is that it costs twice as much for RAID-1 vs. single drive. But at least then you've got some protection. FWIW I use a mirrored pair of 3.84 TB Samsung PM863a in my daily driver workstation, plus an old WD RE4 spinning drive for regular backups.

The question is what checks regularly that no bit was flipped in data?

[BillBroadley]

I did some research, the sn850 has a pretty large SLC cache (280GB or so), the samsung 980 pro has a small SLC cache (about half of the sn850), and both are slower than the 970 pro (which didn't have a cache). Not sure about the 980, other than it doesn't have a ram cache.

Depends on your use case, sure some use cases will exceed the cache size, and it's disappointing that the out of cache performance is less than the previous generation. But then again in most normal desktop/laptop use cases these new drives rock.

[torsionbar28]

Agreed. And I am still scared of buying an SSD, feeling that it will fail all of a sudden (unlike HDD which fails gradually).

I made the switch only last year. The key is to use mdadm (or ZoL) and create a mirror. Plus regular backups to some external media. Downside of course, is that it costs twice as much for RAID-1 vs. single drive. But at least then you've got some protection. FWIW I use a mirrored pair of 3.84 TB Samsung PM863a in my daily driver workstation, plus an old WD RE4 spinning drive for regular backups.

[spirit]

The key to SSDs lasting is not filling them all the way up....if your SSD is at under 75% full it should last plenty long, if you get over that you are reducing the lifespan of the SSD even on models that are have built in over provisioning as there are fewer cells to spread the load over.

If you plan on having 200GB of data, don't buy a 256... buy the 512 etc...

or buy a datacenter ssd, where you already have the spare (1,6TB instead 2TB, cause we have 400GB spare). + you'll have a supercapacitor and bigger endurance. I'm filling my datacenter ssd/nvme around 90/95% without any problem with a lot of constant write.