Current-Users archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: SSD



>>>>> "ms" == Michal Suchanek <hramrach%centrum.cz@localhost> writes:

    ms> Slow write performance tends to be the major factor behind
    ms> slow adoption of SSDs

People on the ZFS list say the SSD's with SATA interface are always
faster in io/s (like 10 - 50x faster) compared to single spindles of
high-RPM SAS, and some of them are faster for streaming write
throughput as well.

    ms> the datasheet says that the disk has dynamic wear leveling

they all have this.  no exceptions.  FLASH cannot work without it.

If it's SSD with SATA interface, the wear-leveling issue is
fragmentation:

 http://www.pcper.com/article.php?aid=691

which silently degrades performance.  Whether or not it happens
depends on the quality of the closed-source wear leveling algorithm
they use.

I have this in my notes about SATA SSD:

-----8<-----
There's a problem in the MLC SSD drives with particular JMicron
controller chips where some patterns of small writes trigger ~~10sec.
hangs, which is most often called "stuttering" in online discussions
of it that I've seen. My particular example is an OCZ Core 2 64GB
unit. (David Dyer-Bennet)
-----8<-----

If the problem is real, it's likely not just OCZ.  There were other
rumors that Mtron and Imation are the same, and that all the Dell
SSD's are Samsung.  I don't know how many genuinely unique brands are
actually in play right now.

and MLC vs SLC used to be something about which people dragged out
whiteboards and ranted for hours, but it is now a little silly because
most drives marked MLC have a small SLC write cache in front, and
sometimes also DRAM+supercap.  It seems like it's really the
proprietary control software that they're selling, not a commodity
microchip plus board, so figuring out how many unique brands are in
play is what needs doing now, not last year's MLC vs SLC.


However, if by SSD you mean ``USB stick'', yes, writes are
mind-blowingly slow.  Some of them also have really crappy wear
leveling and won't outlast a month if you are doing builds on them.
Others do fine for much longer.  The price of good sticks vs bad ones
doesn't differ.  Have a look to the reviews on newegg---that's the
only way I've foudn to tell them apart.  Then I just bought four
different brands and waited to see which ones survived.  

Finally understand that NAND FLASH has a read-disturb property, so the
wear leveling algorithm needs to keep track of how many times a cel is
read, and rewrite it before it gets marginal---how much you want to
bet USB sticks just skip this part?

Anyway, of the ones that die within a month, some fail totally,
meaning they don't even attach any more.  Others are sneakier and fail
by becoming read-only.  I had one in a Linux box with as much RAM as
the used space on the filesystem, and I didn't notice it had broken
several days past until weird behavior made me look in dmesg, so even
though the thing was ``clever'' enough to preserve my already-written
data by failing to read only, the crappy exception handling of the
overall storage stack meant that I still lost several days worth of
writes that never made it to disk and never returned an error to
userland.  The storage failure handling of most OS's seems to really
blow chunks.

Attachment: pgpAVlnbQLQvD.pgp
Description: PGP signature



Home | Main Index | Thread Index | Old Index