Solved Optane M.2 nvme 4k format failing with error

We did learn the timeouts in nvmecontrol are too short for the format command so seems like it is worth ensuring those can be extended for successful formats as might be needed now and then. I'd expect secure erase operations would also be fairly slow, even with fast devices, unless it is just flushing a key or something.
 
I wonder how nda(4) driver compares to nvd(4) NVMe driver that you are using.
-CURRENT uses nda(4) as default now. So that is the future.

nvmecontrol devlist versus camcontrol devlist is a difference.
So flashing may need adjusting in the future.
 
Certainly it may be worth revisiting in the future, and it won't be nearly the hassle it was this time once the timeout complications are resolved.

Performance between formatting 512 and 4k LBA formats is clearly deep in the noise floor. For SLOG (sync random writes) 512 is possibly slightly faster, but the difference is well within the noise. Claims of +10% performance from 4k over 512 were not realized with this device (SSDPEL1K100GA, 100GB Optane™ SSD DC P4801X M.2).

LBA/FW ver.4k E2010650512 E20106504k E2010485512 E2010600
Median Mb/sec759.20762.30757.50742.80
Average Mb/sec721.70722.87721.64724.35


sync_random_writes_optane_512v4k.png

Side note, this is one of the things that I love about the BSD community: had the formatting bug had been a windows kernel bug, it would be "won't fix" as it isn't commercially relevant given Optane is discontinued; with FreeBSD we had a kernel level resolution validated in less than 24 hours.
 
Claims of "10% performance improvement" never materialize :)

See all the Intel, AMD, and Apple Silicon claims before the last CPU releases. It is actually quite annoying.
 
Erichans said:
Your reported synchronous write results (512 or 4K) seem quite in line with Intel Optane DC P4801X Review 100GB M.2 NVMe SSD Log Option

Wonder under what circumstances these claimed +10% gains have been measured ...
Click to expand...
Most likely someone somewhere tested a particular case and it became received wisdom loop cited around an ever widening circle seemingly growing ever more credible by the number of citations discoverable. Though the firmware reporting "Good" and "Best" performance for 512 or 4k LBA formats certainly would seem to validate the premise. There's likely some basis for it to have gotten into intel's annotations other than a random blog post, but who knows.

Anyhoo, the little DC sticks with supercaps are pretty awesome cache devices. It feels a bit like the death of the Concorde or the SR-71: the future died a little. Hopefully, some entity with a bit more marketing savvy than Intel will bridge the void eventually.

I've got 10x 800-GB 2.5 SAS 12G SSDs configured as raidz2 and a bit slowed by aes-256gcm/blake3 and compression=on, but that array yields (quick test) write: IOPS=7351, BW=459MiB/s.

The optane mirror, not striped, yields write: IOPS=10.4k, BW=652MiB/s with compression=on

a tmpfs RAM disk yields write: IOPS=21.1k, BW=1,318MiB/s

So... yah, optane FTW. It may be a long time before we have access to any new storage tech quite so flexibly awesome.
 
Yah, I looked into them for my server. It's really too bad that they're discontinued as well. I really do hope someone takes up the quest. It's kinda tragic that Intel basically solved NV storage as well as anyone could, nearly infinite endurance, extraordinary write speed, low power and vastly cheaper than SRAM ($5k/gb). Maybe FRAM will fill the void before it is time to, once again, replace the hardware.
 
You must log in or register to reply here.
Back
Top