On 5/23/2010 11:30 PM, Fred Liu wrote:
Hi,
I have hit the synchronous NFS writing wall just like many people do.
There also have lots of discussion about the solutions here.
I want to post all of my exploring fighting done recently to discuss
and share:
1): using the normal SATA-SSDs(intel/ocz) as ZIL device. For intel
just EOLed
50nm SSDs product line and the spec of the x25-M G2(34nm) 160G is
getting
more decent, I used this one. I get good performance boost
compared with no slog.
I also manually did unplugging power server times, the outcome
seems good -- I got
no data corruption. But there is always potential risk -- does it
have built-in ram cache?
We get no clear answer from intel.
yes, both the X25-M (both G1 and G2) plus the X25-E have a DRAM buffer
on the controller, and neither has a supercapacitor (or other battery)
to back it up, so there is the potential for data loss (but /not/ data
corruption) in a power-loss scenario.
Sadly, we're pretty much at the point where no current retail-available
SSD has battery backup for it's on-controller DRAM cache (and, they
/all/ use DRAM caches).
2): using PCIE-SSD(i.e. fusionio's ioDrive). For the universal
standard of PCIE-SSD is still on the way, the responding driver
under solaris become apparent. And also the cost is very high.
Yes, you do have to find those with Solaris drivers. Fortunately, pretty
much all the manufacturers recognize the big market that ZFS/Solaris has
for these devices, so most already have Solaris drivers, and the rest
are almost universally working to produce them.
3): using PCIE-DRAM-with(out)-SSD(i.e. marvell's write acceleration
module (WAM), ddrdrive X1, curtisssd's
HyperCache. It has the higest spec but the same driver and high
cost issue.
Yup.
4): using SATA--DRAM-with(out)-SSD(i.e acard's 9010, curtisssd's
HyperHD). No driver issue and price is in the middle.
I also tried acard 9010, the result is good enough.
Frankly, these are about the best available solution, without breaking
the bank. The bad news is that they're not quite read-for-prime-time yet
in terms of support and packaging.
So it seems we have to wait awhile to get a really ideal ZIL device
solution. If intel/ocz can confirm no ram cache, that will be good news.
Thanks.
Fred
Nope, ALL SSDs use DRAM caches for their controllers. So far, I'm only
aware that the Zeus stuff (plus a couple of industrial/military-only
products) have battery backup for their controllers. Idiots - it's like
$0.50 in parts, and a one-time $10,000 in engineer design/qa time to put
the damned thing on, and they could charge a 25% premium and we'd pay it.
Frankly, I'm really surprised that there's no solution, given that the
*amount* of NVRAM needed for ZIL (or similar usage) is really quite
small. a dozen GB is more than sufficient, and really, most systems do
fine with just a couple of GB (3-4 or so). Producing a small,
DRAM-based device in a 3.5" HD form-factor with built-in battery
shouldn't be hard, and I'm kinda flabberghasted nobody is doing it.
Well, at least in the sub-$1000 category. I mean, it's 2 SODIMMs, a
AAA-NiCad battery, a PCI-E->DDR2 memory controller, a PCI-E to SATA6Gbps
controller, and that's it.
--
Erik Trimble
Java System Support
Mailstop: usca22-123
Phone: x17195
Santa Clara, CA
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
