XP256 - Mirroring

[Subbu Krishna]

Is there any reason why someone would do a RAID5 and mirror all theRAID 5 disks on
a XP256. Any performance / cost issues. Would it be more prudent to do a RAID5 or
RAID-1 instead of doing both.

Thanks for your input.

SK

[Mike McKinlay]

RAID5 mirrored? Either your system/data is really, really, really important, or it was done because it could be.

True RAID5 can stand a single drive failure without loss of data. However, if the array card fails, then the drives are unavailable.

I suppose a mirrored RAID5 array could withstand an array card failure, but only if the second RAID5 array was on a separate array controller. The mirroring would also need to be supported separately.

[Victor BERRIDGE]

The only reason I see is, 24h/7days systems where you have no money problem (cost).
It can be for backup purpose =>
you mirror then you split and backup the copy then synchronize again...
Is it worth the trouble up to you to decide...
Because for crash recovery on an XP256 / HDS770E
You can do that internal using shadow image
...
I know of cases where ther is mirror copies of arrays between 2 distant cabinets, in case of system failure the distant host can replace the one down...
I dont believe what Mike sais avout the controller since evering thing is redudndent on these bays, but that also mean you have 2 controlers on your host also (alternate links)
For me RAID5 on these bays is already very very safe indead

[Dave Wherry]

As Mike and Victor said, it can be done. I just can't think of any reason you would want to.
On an XP256 a Raid-5 marketing group consists of 3 data and 1 parity disk (normal Raid-5 stuff). If you lose one disk you will continue to work. No need for a mirror. For that matter an XP should have a hot spare disk in it and that works just fine. I the spare work here 2 weeks ago when a CE accidentally knocked the terminator off of one of the disks. The hot spare activated and users never noticed a thing.
The fewest number of host ports you can have on an XP is 4. You buy CHP cards in pairs and each FC-AL card has 2 ports and SCSI cards have 4 ports each. Since you have the multiple ports just use alternate links and you do not have to worry about a controller on the host or XP failing. Works seemlessly.
If you are using Mirror/UX the host is doing 2 writes, 1 to the primary and 1 to the mirror. No need to put that extra I/O on your system. The Raid-5 has already given you the redundancy you need.
As was said you could do a mirror and split it off for backups. You've already made the investment in the XP, use Business Copy. It is much faster and more flexible.
If you did have a 2nd XP in a remote/disaster location you would need Continuous Access (CA) to even see those disks from your local host. CA essentially does the mirroring for you. Again, why have the host do 2 writes for each I/O? Let CA do its' job.
As you might be able to tell I really like my XP. You just need to make a disconnect from the old JBOD world. So many of the "disk techniques" that have been taken as gospel over the years really do not apply in an IDC array world.
I worked in another shop a couple of years ago and the Senior Architect evaluated an EMC Symmetrix and rejected it mainly because he could not accept some of these new concepts. It's really too bad because and IDC array would have made a world of difference in performance and managability.

[Tim Malnati]

My familiarity with XP equipment is minimal, but my understanding is that it is very similar to EMC equipment. My thinking is that some configuration downtime issues may have caused this to happen. The original raid 5 defined drives may have been performing very poorly due to other activity on the same physical platters. A decision may have been made to mirror them to another set in an attempt to improve read response time. Not a great way of handling things, but maybe the only alternative available at the time.

[Victor BERRIDGE]

There is a philosophical difference :
EMC claims they are the best...( But when it comes to compare with HDS they argue that it is true in their configurations mirroring...)
So EMC flavors and is optimized for RAID0-1
HDS is optimized for RAID5
So XP XXX should perform best in RAID5

[Tim Malnati]

Victor,

I just have a lot of trouble with the idea that Raid 5 can ever outperform Raid 0/1. There is a lot more overhead with Raid 5 and the physics involved complicate it further. I would love to see the actual statement that is making the claim. If HP could actually achieve better performance with Raid 5 vs 0/1 there would be no reason to continue manufacturing the 12H array in its current configuration. This is not to suggest that the XP is not as good an EMC, I just can't agree with the claim.

[Dave Wherry]

Tim,
I've emailed back and forth with an engineer at HDS and he has sent me a white paper showing that in their tests in an Oracle environment, Raid-5 outperforms Raid-1. I've recently moved all of my data from Raid-1 to Raid-5 and have seen a minor performance hit. The reason I've seen this minor hit is I added new drives and have not yet finished balancing the data across the drives. I have all of my data concentrated on fewer drives right now. When I finish I expect performance to be as good as Raid-1.

Raid-5 usually performs worse than Raid-1 when doing writes. When using an IDC like the XP or an EMC, all writes go to cache. There is no performance penalty using Raid-5.

The XP and EMC really change the way you have to think about disks.

[Tim Malnati]

I have performed a full suite of benchmark tests on EMC equipment with results that are far different. These tests involved real data using everyday processing. I fully agree that the size of the hardware cache will drastically impact write performance. The size of this cache will also significantly impact sequential reads. But when you get into random read situations, the hardware cache helps some, but often becomes a situation where calls are being satisfied directly from disk. Raid 1 is clearly better than Raid S in this situation. Theoretically Raid S should not suffer as much as it does, but the benchmarks say otherwise. We could get into the age old differences between Raid S and Raid 5, but I will leave that for another day. The situation when you add striping to Raid 1 to make it 0/1 was amazing. Throughput increased to levels that Raid 1 or Raid S could not hold a candle to. How the stripes are spread within the EMC is critical where you need to balance things out among all the controllers in the unit. Balancing SCSI connections is also important. We also tried striping Raid S and got the poor results we expected. Another important factor to keep in mind is the size of the memory buffer cache in the connected server. Keeping this buffer optimized makes all the difference in the world when you are dealing with multiple machines attached to the large storage units. When we compared the overall environments with an optimized EMC group using 0/1 with another EMC using Raid S (using the same application), management made sure the money got into the budget to move things toward the 0/1 arrangement. Things were better at the machine level also, a 4 way K460 was outrunning a 6 way T600 with a near identical environment otherwise.

Don't try to read into this that I favor EMC. I have not had a similar opportunity to play with XP gear, but I would not be surprised if I got similar results in this environment as well.

[Dave Wherry]

Tim,
I agree with your points and you've agreed with some of mine. What this shows is there are a bunch of variables to consider and those variables change in different environments. There's not a blanket answer.
So many times we see replies in the forums that just say this is the answer and that's it! Those either aggravate me or crack me up. Usually depends on how much sleep I had the night before.
So people will continue to agree and disagree. It's good to see the different viewpoints, although some of that is just speculation. It's great to see the experiences people have had. Especially like the benchmarks you've done.
Regarding the original question here, striping Raid-5 just seems silly.
Have a good holiday.