Yes, although I’m pretty sure that is not the whole story. The persistent write cache’s NV memory that the 512e MG08 hdd cache will dump to is 4MiB (two W25P16’s), substantially more than the 4KiB required to buffer a sector for another go around because of RMW. - I’m 80% sure those were the chips on the PCB, its been awhile since I’ve seen them.
I know for a fact the 512e x300’s had 3MiB divvied up amongst 3 NV memory chips for it’s persistent cache.
The actual PCBs between 512e and 4kn drives are different, the 512e (PWC) drives have more expensive PCBs that are of a different design and populated with more components to facilitate the PWC.
I got the MG08ACA16TE. SATA. Toshiba website states that there is a 4kn version, but I’ve never seen that model in retail ever. And my SKU has that persistent-write-cache, for better or for worse. Might get me into trouble in a year or two. I heard it might not work well with Ceph.
I’m not sure why 4kn sata drives seem to be so rare, maybe there are a controller compatibility issues like there was for SAS? although that was solved many years ago.
My understanding of the write caching enablement issue with ceph is that when write cache is enabled at the osd level it keeps sending flush commands to the harddrive which may or may not honor the request. For hdds that honor the request explicitly, iops tanks, but some hdd manufactures will claim a write complete even when it hasn’t and still maintain good performance throughout the flush commands it receives.
Disabling write caching just stops the flush commands, the actual hardware level write cache on the drives is still used.
Yes, I believe all the modern exos do… but do all sata controllers?
the possible lack of sata controller support was the only reason I could think of for why “out of the box” 4kn sata drives seem so rare.
I get the feeling the way exos are doing the 512e>4kn transition is kind of a half-baked approach to implementing lba sizes “correctly”, I would assume it would take a firmware update to change the way the buffer and elevator are handled between the two lba sizes to optimize performance.0
— which makes the Seagate ST18000NM000J results all the more strange since it shipped in a 512e configuration but is performing better with 4kn lbas
Btw, I have some old WD drives. They are also Advanced Format drives.
Since I’m reorganizing all my data I figured I might look into 4K-ing those too.
I have emailed WD support in the hopes they provide me with the relevant vendor-specific utility to switch them to 4K.
If they do, I will perform the test above with at least one of them as well.
P.S.
I actually did a firmware update along with the 4K switch. From SN02 to SN04.
Amusingly enough the firmware document explicitly states the new version was not compatible with 4K. Been working perfect so far.
The two Exos X18 drives I just installed shipped 512e but were convertible to 4kn in ~50 seconds (after 60 seconds of warnings that the process may “render the drive unusable” ) using SeaChest. Both modes are supported by my system’s C622 chipset SATA controller and VROC RAID. Performance with Crystal Disk Mark is identical in both modes.
I tried using hugo, but it says my devices are not supported. @Mach3.2 What drives do you have where it worked?
it is strongly recommended to contact Western Digital and provide your hard drive serial number, to make sure you obtain a compatible low-level maintenance tool.
For now I’m following this advice. Hopefully they will get back to me with a version that works.
I think my drives are simply too old to be supported.
My life story
I actually have a pair of 750 enterprise drives from 2009 in my machine. In raid 0. Been raided since then. I think they’re still 512n drives.
The other WD ones are a 2TB and 3TB drives and they’re 4K drives (512e).
The 2TB can’t even hit 100 MB/sec. Not that it’s ever needed to. Been my media drive for probably a decade now. MBR must still have been the default when I got it, because I just reformated it today. (Tho it is 512e/4K.)
At least the 3TB one can saturate a gigabit link. Barely.
I ran CrystalDiskMark - turns out the oldest, raided drives are STILL the most highly performing out of the entire bunch. They’re not even SATA3! (Excluding the SSDs ofcourse. The PCIE4 nvme drive I got the other day blows every other drive I have COMBINED out of the water - including the SATA ssds. )
I remember I was so happy to get them for my oldest build. I had done a little bit of research back then, and wanted a lot of performance, so my rationale was, I will get enterprise-grade drives that are built to be reliable so that I can RAID-0 them and have my cake and eat it too. High-school me was so innocent.
But I’ll be damned, cuz they’re still going strong. I promise to run those drives until they come off the hinges.
I’ve been researching this whole topic for a while now and although I have learned a lot, I still feel like I’m not quite there yet when it comes to actually placing an order.
these 512e/4KN Fast Format drives appear to only be able to support 512, 520 and 528 and nothing in the 4096 range. So I feel like I’m missing something here.
why can they support all kinds of formats 512 onwards and 4096 onwards, while all other Exos drives seem to be sold in specific “configurations” i.e. either 512 or 4096.
These drives all have 4k native (4kN) sectors but present 512 blocks (512e stands for 512 emulated). So at the lowest level they are 4k all along.
These other sector sizes are a SCSI/SAS thing. The extra 8 bytes of the 520 are for error correction. I‘d assume 528, … are the same but even more redundancy
I did write Z80 assembly on CP/M to for hardware support, but I had no HDD at the time. That role was taken by a giant 720k floppy, which felt like a hard disk compared to the 160k primary Apple ][ drives (CP/M ran on a Microsoft Softcard).
My first own HDD was a 20MB unit on an 80286 PC and yes, it also used 1:3 “interleave” (which is what that skew was called) so the controller could catch up, and floppy type stepper motor for head positioning: voice coil (sleds first, swivle came after) and reduced average seek times from 90ms to 30.
On the 80286 it had gone to PIO mode, because that was way faster than the 8-bit wide DMA used on the original 8088 based PC, which was extended to 20 bit phyiscal addresses with a separate chip for the higher bits. With that 8-Bit DMA the interleave was more like 6… and only until you had to reprogram DMA and shift registers it for a different segment (well that was all in the BIOS, for which IBM sold a big book with the source code).
And then came the first RLL controllers that not only gave you 1/3 more capacity using a better physical bit encoding than FM or MFM, but also faster chips so the interleave could be reduced to 1… So by using a different disk controller the same physical disk got 1/3 more capacity and 4x the sequential speed, but sector size and seek times obviously couldn’t change.
So you had disk optimizers running all night, to ensure that all your files were nicely layed sequentially all across your dozens of megabytes of drive space…
But with 640k of (parity) RAM and a System V.2 Unix from Microport fitting on 3 floppies files were much smaller, too.
) using SeaChest. Both modes are supported by my system’s C622 chipset SATA controller and VROC RAID. Performance with Crystal Disk Mark is identical in both modes.
Not that it’s ever needed to. Been my media drive for probably a decade now. MBR must still have been the default when I got it, because I just reformated it today. (Tho it is 512e/4K.)
)
