Mirror disk or Raid 5

[Steven Chen_1]

Hi,

I would like to seek advise about Oracle on HP L series platform, partically those of whether mirrored disks or strip disks (raid 5 using va7100) is good for Oracle 8i.

Thanks for comments first.

Steven

[Jim Turner · Accepted solution]

Steven,

If you can afford it, always go with RAID 0/1 ("striped mirroring" for lack of a better description). This gives you the highest throughput and resiliency. You'll be out more money and rack space, but it is the fastest way to go.

The downside to traditional RAID 5 is write performance. Specifically, most RAID 5 schemes require a partial RAID stripe to be read from disk back to cache where it is completed and then written back out to disk. Less money, fewer disks, and less rack space, but more potential for lower performance.

AFAIK, the only HP storage product that overcomes the write-performance hit on RAID5 is the XP256/512 (SureStore E Disk Array XP) where partial RAID stripes are *always* held in cache and only flushed to disk once they are complete. The VA7x00 product may do the same thing, but I am not familiar enough with it to say.

Either way, you should be in good shape with speedy new kit. The only problem I've had with my Data Warehouse (N4000 w/XP256) has been APs using horribly inefficient SQL and blaming it on my system tuning. Once I convinced them to rewrite the garbage SQL that PowerMart vomited out, a multi-day process dropped to a couple of hours. Imagine that ;-)

Cheers,
Jim

[Bill Hassell]

RAID 5's major advantage is that it requires only 5 disks to protect 4 disks' worth of data and reconstruction of a broken disk is possible for any disk in the group. Mirroring means that 8 disks are required to protect 4 disks' worth of data.

RAID 5 requires a lot of work to implement, essentially coordinating the computation of the parity information and writing of all the stripes at the same time. Normally, a hardware controller will do all this work. Starting at version 11.11 (11i), the VxVM volume manager can perform this task in the CPU. However, you will be trading the cost of a disk array controller for HP-UX overhead.

In contrast, mirroring can be handled in hardware disk arrays as well as software in the CPU quite rapidly. In fact, with standard LVM mirroring, multiple paths to the disks can be load balanced for reads automatically. The only downside is the number of disks. But with 10's of Gbyte disks going for less than $1000 US, mirroring should be a minimum for production systems.

[Don Bentz]

Perhaps somebody more knowledgable can comment on this, but if you are using a VA7100, I believe that is an AutoRAID device. You don't have the decision to make regarding the layout except when it comes to deciding how much unallocated space there will be and by implication, the amount of RAID 0/1 space is reserved in the device.

[Volker Borowski]

Hey,

may be I am old fashioned, but I do not like autoraid so much. I like to know exactly, that those parts I need for recovery are stored on sepearate physical units, whatever RAID each of these components have.

Autoraid is cheap, a good filesystem storage, a good database-datafile storage, but I would never mix DB-data with any LOG- or controlfiles in a single autoraid box.

Autoraid and the ease of using and configuring it, lead to less experienced people doing configuration work, which they better should not do. Somehow it is the same with LVM, if you extend filesystems in SAM... and *whoops*, /ora_data1 shares the same physical disk as /ora_arch.
You can avoid it, if you do it manually.
But with autoraid, you can not, even if you would have known better...

Considering performance and assuming only datafiles on this autoraid-thing, it is really a wonder. If I understood the manual correctly, it is profiling the io-behavior on each logical unit, and decides, if it has to reorganize from RAID1 to RAID5 or back, as long as space permitts. This is really amazing!
I do not know, if it is able to profile, that you better do not put data and coresponding index onto the same disk for performance.

Conclusion:
Good and cheap storage, but you have to use it properly.

Volker

[Don Bentz]

I agree with Victor. The only way I'd use an AutoRAID (again) for a database is if I had separate AutoRAID boxes for the DATA/INDICES/TEMPORARY TABLESPACES/ROLLBACK SEGMENT/REDO LOGS components. The administration is easy, but I don't want requests for data waiting on completion of a request for an index or write of a rollback segment on the same spindle. But that's just my opinion.

[Steven Chen_1]

If Oracle 'lives' better in mirrored disks (suggested by Oracle), do we still need VA7100?

[Vincent Fleming]

One thing you guys are missing... the VA series uses RAID-5DP, not RAID-5.

What's that, you ask... Double Parity. It uses two parity schemes and has 2 parity blocks for each redundency group. See the attached white paper for details on how it works. In a nutshell, the benefit is better fault tolerance - the array can suistain a failure of ANY TWO drives, and continue operation. RAID-0/1 can't do that.

Let it suffice to say that the write penalty for RAID-5DP is probably a bit more than RAID-5.

Now that said, you're also forgetting that a VA 7100 has 2GB of cache. That's a LOT of cache for a box with 15 drives!!!

The cache will insulate you from most performance problems. The AutoRAID automatically moves data around in the array to optimize performance so you don't have to be concerned about if the data and logs are on the same drives.

The VA also allows you to fix it in RAID-0/1 mode for best performance, so you can tweak it manually.

AutoRAID has taken some major improvements in the VA. Give it a try.

[Dragan Krnic]

You don't.

Bill Hassel, it's a good description of what 5 is doing and why it is so slow (except possibly on XP's). But your argument that having 5 disks protect the net worth of 4 at a lower price is a bit too light because you forget that a RAID controller itself cost usually more than 3 disks it saves.

My experience is that nothing beats S/W mirroring for both price and performance.

[Dragan Krnic]

And for all of you lovers of separate physical media for tables, indices, logs and traces: If you take a little effort to just spread a file system over sufficiently many spindles you get the separation for free. No sweat.