PM983 or other 8tb ssd for sustained writes?

[Aardvark0]

Looking for ssds with sustained writes that don't drop to super slow when the cache runs out and drive fills up.
application is to dump 1tb or so continuous to the drive in one go.
source maxs out at 2 GB/sec so gen4 is not necessary.
currently use EVO 860 sata ssds, can fill the entire drive at a sustained .5 GB/sec
the 8tb m.2 consumer drives available now are qlc - corsair, sabrent rocket, mushkin alpha - drop below this on sustained writes exceeding 100 GB as the drive fills.
SN750 is the only 4tb nvme that I've found that offer better sustained writes, everything else is only 2tb - evo 970/980, sn850

looking at u.2 format , PM983 seem to sell at a decent price on ebay in 7.68 size, and is rated at 1.9 GB/s writes, does that translate to sustained writes since its an enterprise ssd? can the firmware be updated since these are oem? these have decent TBW, but does samsung honor the oem warranty?

the Hynix PE6011 is also reasonable but does not seem very popular.

 

[EffrafaxOfWug]

I've not used the large versions myself (only the 2TB M2 version) but the Micron 7300 Pro's are usually quite a lot cheaper than the Samsung's and being enterprise drives can sustain the 1.8GB/s writes. The 7300 Max sits in a slightly higher performance/endurance ballpark.

If only the sequential speed is of a real concern, is RAIDing some smaller drives an option? If you've got a spare 16x PCIe slot, four M2s in a bifurcation card are a popular way of creating large fast drives, and I've got an array of large SATA SSDs on an ancient 9211-8i that's still capable of silly speeds. Or are you trying to keep it simple with just a single drive?

 

[MBastian]

I've not used the large versions myself (only the 2TB M2 version) but the Micron 7300 Pro's are usually quite a lot cheaper than the Samsung's and being enterprise drives can sustain the 1.8GB/s writes. The 7300 Max sits in a slightly higher performance/endurance ballpark.

Official specs say they max out at 1GB/s writes. Probably the "you'll get at least that, now matter what" value.

If only the sequential speed is of a real concern, is RAIDing some smaller drives an option? If you've got a spare 16x PCIe slot, four M2s in a bifurcation card are a popular way of creating large fast drives, and I've got an array of large SATA SSDs on an ancient 9211-8i that's still capable of silly speeds. Or are you trying to keep it simple with just a single drive?

Had the same question. You can hit <Dark Helmet>ludicrous speed</Dark Helmet> in a RAID setup.

 

[Evan]

PM983 is ok, it’s a lower power drive, 2GB/sec is right at its limit I would say, I have some 3.84tb versions but HPE firmware so maybe different that your looking at.
If you want max performance the it’s going to be in a 15mm and 25w kind of format rather than 7mm and 10w or so max aimed for balanced performance. Something like the 6.4tb HGST SN840

 

[EffrafaxOfWug]

Official specs say they max out at 1GB/s writes. Probably the "you'll get at least that, now matter what" value.

Odd, you're right - store I usually buy from has it clocked at 1.8-1.9GB/s though so there's a obvious discrepancy there. My 2TB 7300MSs are certainly faster than the rated specs but I've not tried a sustained seq write torture test. Incidentally, when I did do a "light" torture test (three drive writes) on an Intel P4101 that was able to exceed its supposed 650MB write speed by ~30%.

If it needs to be >2GB/s seq write then "enterprise-lite" on a single drive probably isn't going to cut it and you'll need to look at the more hardcore series like the Intel P4xxx's, the Micron 9300s or the Samsung P17xx's, all in 2.5" U2 format. Models in the 8TB ballpark can usually write at 3-4GB/s. These can get expensive quite quickly so you can always think of RAIDed SATA/SAS/M2 as a potentially cheaper option.

If the original dataset is "only" 1TB, is the 8TB drive a nice-to-have or a requirement for collecting several datasets in short order?

 

[UhClem]

...
currently use EVO 860 sata ssds, can fill the entire drive at a sustained .5 GB/sec
...

Can you do that twice ? Or, only (once) directly following a (full-drive/secure) erase ?
Just curious ...

 

[Aardvark0]

Thanks for the suggestions
I have (4) of the sn850 gen4s in a raid for testing - its an impressive 21 GB/s read, 18 GB/s write for the entire 8tb.
I can't raid for this particular use. I can get the micron 9300 or the pm983 7.68tb for about $1k which seems ok. Do you know if there is a util to format the pm983 or the 9300 as 4kn?

with the 860 evos, I use until full or almost full, then just quick windows reformat, usually transferring between .5tb and 2tb at a time, don't see any slowdowns. maybe if your deleting files you would see a performance hit.

 

[EffrafaxOfWug]

Do you know if there is a util to format the pm983 or the 9300 as 4kn?

msecli has an option to set the formatted block size although I don't know if the 9300's support 4kn.

root@wug:~# msecli -J -?
USAGE:
msecli -J [ -b <NumBlocks> -l <Block Size> ] [ -p <Protection Type> [-e <Protection Interval Exponent>]] [ -d ] [ -c ] [ -f ] -n <device-name>

    Performs Format Unit operation on a specified SAS device.

-J
    Performs Format Unit operation on a specified SAS device.


    -b <NumBlocks> -l <Block Size>
        Suboption of '-J'. This option used to resize the device to
        the number of blocks specified with -b and using logical block
        size specified with -l.

    -p <Protection Type>
        Suboption of '-J'.This option used to specify the protection type to use.
        Valid options are 0-2.

    -e <Protection Interval Exponent>
        Suboption of '-J'. This option used to specify the protection interval
        exponent for the desired protection type. This option can be specified
        with -p and only valid if protection type is set to 2.

    -d
        Suboption of '-J'. This option used to set the Disable Primary bit.
        If set device will not use the PLIST (Manufacture bad block list)
        to identify detects. Default is to not set this which results in
        the device not assigning LBAs to parts of the device identified
        as defective in the PLIST.

    -c
        Suboption of '-J'. This option used to set the Disable Certification bit.
        If set device will not perform any medium certification or format
        verification. Default is to not set this which results in the device
        performing medium certification and adding any detected defective
        areas to the GLIST (Grown bad block list).

    -f
        Suboption of '-J'. This option used to set the Stop Format bit.
        If set the device will terminate the format operation should it
        encounter any list locate or access errors. Default is to not set
        this which results in the device completing the format operation
        should it encounter list locator or access errors then return check
        condition for the format unit command.

    -n <device-name>
        Mandatory suboption of '-J'. The operation is performed only
        for the specified SAS drive <device-name>.

    -r
        Global optional parameter. Signifies silent mode and suppresses
        the command output.

    -s <out-filename>
        Global optional parameter. If <out-filename> is specified, then the
        command output will be logged into the specified file.

Copyright (C) 2020 Micron Technology, Inc.