I have a REALITY environment installed on a RAID 10 environment, and
use RAID 10 on all of my other servers, as well. I'm about to move
all but REALITY (at least for the time being) to Virtualization. The
proposed server has 16 drive bays. Proposed solution by vendor is to
use 4 x 300GB 10k RPM drives, using RAID 5. My experience for so
many years has been with RAID 10, and everything I read, basically,
says that RAID 5 can/will be slower in response to "writes".

I have asked to increase the number of drives, since the bays are
available, and reduce the capacity of each drive (e.g., instead of 4 x
300GB, go to 6 x 200GB), or some such arrangement. The change would
require 12 bays to be occupied (for RAID 1), but would allow for more
efficient use of RAID 0 (striping). And, yes, I realize there is
more cost involved, but my major concern isn't so much for "what are
we paying", rather, "what's the best method to protect our
invenstment".

Does anyone have any experience with virtualization, be it with MV or
any other arrangement?

Thanks, in advance,
Jim Cronin
DIR. MIS, Kittery Trading Post

On Feb 9, 8:22*am, JJCSR <JCro... (AT) ktp (DOT) com> wrote:
I'm a bit confused. A fair bit of the post pertains to RAID
configurations but it appears the base question is in regards to
virtualization. I can't speak to the performance issues in regards to
number of drives and RAID config but we do have some experience with
virtualized d3 servers. We have a number of customers running Win200x
server and either Hyper-V or VMWare virtuals with d3/linux. I know it
seems silly to use Windows as the server O/S but this is how we avoid
d3/NT problems (spare me the outcry) in places that "must" run MS.
The oldest of these installations is about 2 years now and have had no
problems. Actually, we did have a problem where one of the customer's
IT guys put a snapshot from an earlier time in place and all data from
the last 3 months "disappeared". Mr Peabody threw the switch on the
wayback machine. They hadn't been moving the backups off the virtual
either. Boy was that a heart stopper until it was figured out and the
current snapshot restored. It's really great if you are installing
the same base software on multiple servers. Set up the first one,
snapshot it and restore the snapshot onto the other servers. The O/S,
db and valued added software are instantly there. All you need to do
is configure it for that location and activate. That has worked well
with no issues. Quick too. There was a fair bit of trepidation the
first time we used this configuration but I have no qualms about doing
them now.

The issue, herein, with RAID involves the use, or non-use, of RAID 0
(striping), along with RAID 1 (redundancy). Years ago, when I first
started using RAID 0 + 1, the folks at PICK Systems had performed
tests to show that striping across 2 disks increased performance by
about 19% (I believe that was the number). Adding more disks to the
array did not necessarily increase, exponentially, but certainly would
increase "by some extent" - the more disks in the array, the greater
the overall efficiency-increase. Additionally, I wanted hot-swap
redundancy, so I had to sacrifice a bit of cost to the array (each
drive needed a "partner" for RAID 1). In the past two years, we have
had 2 drives go bad on our H-P system. However, because of
"redundancy", we simply yanked the bad drive out, inserted the new
one, and let the system caught up with redundancy of data, on its
own. And, there was "0" downtime.

Now, considering Virtualization, I am interested in knowing if I can
expect increased efficiency from a striped array (RAID 0), with
redundant disks (RAID 1). Or, will Virtualization react differently
to this combination of RAID? Again, as stated earlier, everything I
can find by researching the internet, thus far, seems to imply that
RAID 0 creates more overhead on "writes".

I hope I haven't created more confusion.

Jim Cronin

G'day Jim,

It doesn't matter if you are running with virtualized servers or on
bare metal - the characteristics of your RAID array are constant. What
WILL make a difference is the RAID controller being used, and also the
speed of the drives. With a virtualized environment you are likely to
see much more disk activity associated with multiple OS'es and higher
overall CPU utilization

I'd question the 10K drives. If you can't justify $$D, then I'd at
least spring for 15K ... I'm assuming the controller will be
supporting something like 6Gb SAS?

Though I've been an advocate of RAID-10 ('cause you can be doing 2
concurrent reads!), the latest tests we have performed highlighted
that with the right controller, the performance differentiation
between RAID-10 and RAID-50 wasn't as large as you might expect ...
and given that with many workloads the read:write ratio might be
around 10:1, I wouldn't get too hung up on the write delay (and more
disks in a RAID-50 array has more of a performance impact than more
drives with RAID-10 (and with $$D's, RAID-5/50 makes more financial
sense anyway :-)

In terms of virtualization in production environments, we have had
people leveraging virtualization for 5+ years now, and have used
Virtual Server, Hyper-V and VMware. The only environment we have had
issues with has been XEN, though we didn't establish the environment,
and we had played with XEN internally without issue previously, but we
still don't have anyone running production workloads with XEN at the
moment.

We have done quite extensive performance testing on all of these,
comparing performance to bare metal installs, and importantly also
looking at perfomance characteristics with heavy workloads running in
multiple simultaneous virtual environments ... If you don't have a
platform preference, running your own tests is likely to be
"illuminating"

the trouble that you would face with 16x drives on Raid-5 is when a drive fails. The redundant data is spread around the drives and when a drive fails, the controller will have to do X+1 reads to recreate the missing data. (Your case, 15 reads from the 15 remaining drives, and the bit of data to recalculate the missing drive's data). As you add drives to raid5, the downside is a lot of work to reconstruct the data.
I like Raid 10.
Other ideas: Does the controller support Raid6? I believe that is a raid5 with 2 sets of parity data, so loosing drive you don't get as much of a penalty.
M2C
Rob

On Feb 10, 11:57*am, x <lucian_p... (AT) yahoo (DOT) com> wrote:
Lucian:

Smart minds run on the same tracks, or something like that. I found
that same article about 4 hours ago, and printed it out for my admin
group to see.

As well, I just received a revised quote from my var, following my
request for him to re-evaluate the proposed RAID and disk drives,
stating, first, his company is now in agreement with me that RAID 10
will deliver the best proformance. Secondly, they have replaced the
4x 300GB 10k RPM drives with 6x 146GB SAS 15k RPM, adding just 6% to
the hardware cost of the proposal. Finally, the twin servers will be
delivered with support for RAID 0, 1, 5, 6, 50, and 60.

I'm satisfied. Thanks to Ross and Rob, also, for their input.

Jim Cronin

Check out the FusionIO IODrives if you're planning on running a Linux
server environment. I've not been able to max the 80GB drive we have
in terms of IOPS or bandwidth. Even on our smallish site, we hit CPU
limitations most of the time but I did manage to get 120K IOPS and
about 800MB/s with some undersized file stress testing under D3. The
smallest drive they have now is 160GB and you can stripe multiple
cards with mdadm to get as much storage as you need. You don't need to
worry about IOPS since striping will spread any load over 90K across
the cards. The drives should be low-level formatted to match the block
size of the FS you're planning on using. In our case we're writing
directly to it and using a 4K block size. I've never seen online
resizing run so quickly. Undersized files and SSELECTS are smokingly
fast compared to some larger RAID10 arrays I've run using SAS drives.
The 160GB cards run around $7000 last time I checked, but if you have
a huge amount of transactions hitting disk then you should contact
Ashwood and see if they'll work with you to do some testing. We did
that and FusionIO agreed to pro-rate a return, though we decided to
keep it and bought another one. Once you write to the drive it
consumes the memory so you will have to pay a "refurbishing fee" to
return a demo unit. Still, it's cheaper than keeping it if it doesn't
fit your needs or budget.

GlenB

On Feb 9, 11:22*am, JJCSR <JCro... (AT) ktp (DOT) com> wrote: