RAID setup question/survey for opinions 3

[rws70]

I am putting Exch 2003 on a 2850 with 6 hard disks (4 - 76 gig and 2-146 gig). (70 users)
I am torn between two raid layouts:

Layout 1
OS and Logs and Swap in separate partitions on 2 - 146 gig drives RAID 1
Database on 4 - 76 gig drive RAID 10

Layout 2
OS and Swap in separate partitions on 2 - 76 gig drives RAID 1
Logs on 2 - 76 gig drives RAID 1
Database on 2 - 146 gig drives RAID 1

In Layout 1 the database has a boost in performance from RAID 10, but the logs share the same hard drive as the OS and Swap and thus lose some performance.

In Layout 2 the database loses some performance as a RAID 1, but the logs gets a boost by being on their own set of drives.

Thanks for your help!

 

[xmsre]

The most user visable aspect of exchange performance is writes to the logs. This is why you want to isolate the log spindles. I would

RAID 1 - OS
RAID 1 - logs
RAID 1 - Database

or

1 drive - OS
1 drive - logs
1 drive - database


From a fault tolerance perspective, the first is the preferred configuration. THe second configuration trades fault tolerance on the OS and log drive for added performance/space on the database drive.

If at all possible, go with the first option.

 

[rws70]

on your second config ... why would I only use three of the six drives?

thanks for your comments!

 

[xmsre]

You wouldn't. That's a typo. 4 disks RAID 10 for DBs.

John

 

[lebisol]

keep in mind that in option 1
'partition' is NOT the same as 'logical drive' in RIAD.

if possible keep them separated by logical drives.
my vote goes for option:
-OS/Exchange [RAID1]
-Storage Gorups/Trans Logs [RAID1]
-Priv and Pub Stores [RAID5 or RAID 10]


...don't foget about stores and 110% free space needed for defrag...shadow copy is very nice feature as well...drives are cheep!

:--------------------------------------:

All around in my home town,
They tryin' to track me down...

 

[rws70]

Hi,
Thanks for sharing your thoughts.

One thing I did not understand ...

-OS/Exchange [RAID1]
-Storage Gorups/Trans Logs [RAID1]
-Priv and Pub Stores [RAID5 or RAID 10]

... doesn't that equal 8 drives? I only have 6 drives.

Many thanks!

 

[lebisol]

You are right....IF you can stick more drives then do it rather sooner than later down the line.
hence
"...don't foget about stores and 110% free space needed for defrag...shadow copy is very nice feature as well...drives are cheep! "

All this depends on how many drives your box can hold.

:--------------------------------------:

All around in my home town,
They tryin' to track me down...

 

[lebisol]

look into PE2900 series...not that much more $ for more robust box

:--------------------------------------:

All around in my home town,
They tryin' to track me down...

 

[58sniper]

-OS/Exchange [RAID1]
-Storage Gorups/Trans Logs [RAID1]
-Priv and Pub Stores [RAID5 or RAID 10]

Except that Stores go inside Storage Groups.
Leave just the TLs on that array.

Pat Richard, MCSE MCSA:Messaging CNA
Microsoft Exchange MVP

 

[lebisol]

Except that Stores go inside Storage Groups."

right LOL I started that snowball by typing fast..would be hard to separate

:--------------------------------------:

All around in my home town,
They tryin' to track me down...

 

[xmsre]

1. Don't put the stores on RAID 5. That should be RAID 1 or RAID 10. The write penalty for RAID 5 is too high. You'll use twice as many spindles as you would using RAID 1 or 10 to reach the same performance level using RAID 5. The MS white paper "Optimizing Storage for Exchange Server 2003" says "In general, Raid-5 does not provide the best trade-off between reliability/availability and performance" for a good reason.

2. If you're running Enterprise, and your operational procedures dictate using move mailbox to a new store rather than running eseutil/isinteg, then you'll require free space equal to the logical size of your largest database. If you have four stores, each a logical size of 25GB, then you'd only require 25GB free. If you run standard/SBS and can't create another store, then you're forced to use eseutil/isinteg and have 110% of the size of your database free. If you have only one 100GB store, then you'd need 110GB free.

3. When you size your storage, make sure there are enough IOPS available to support online maintenance and backups without impacting users. If you use backups that leverage the VSS software provider or other copy on write type snapshot providers, the overhad can be significant.

4. When you build your server, avoid the vendor provided automated build program if it first creates a FAT partition and converst it to NTFS later in the build process. This results in an allocation unit size of 512 bytes which causes excessive transactional overhead on the bott drive (read poor performance) You can check the allocation unit or cluster size of a drive by running checkdisk in readonly mode (the default allocation size (including those partitions created by setup in windows 2000 SP3 or later) is 4K.

 

[lebisol]

Interesting...I was told to ensure RAID1 for trans logs and OS while RAID5 works just as well for stores.
In order to get the most space for the $ while still providing reliable and fast solution(15K rpm on this RAID5) I bought into this theory...
This came from MS Exchange Reps whome I handed my pe2900 specs...for what is worth. Not arguing just sharing what I was advised. (or ill-advised)
I suppose the time will tell on ware and tare of these drives :|.
Thanks for the info xmsre.
All the best!

:--------------------------------------:

All around in my home town,
They tryin' to track me down...

 

[rws70]

Question to xmsre

RE: Vendor created FAT partition.

I am using the Dell Open Manage install disk and it does create a utility partition.

I also read Mircrosoft's "How to Align Exchange I/O with Storage Track Boundaries."

Is this what you are refering to?(see quote below) Also, does this mean you do not use a Utility partition? The nice thing about the Dell install disk is it sets up all the drivers, etc. I was going to try and create my partitions first using diskpart.exe and then use the Dell install disk.

(here is the quote about diskpart.exe)
"With a physical disk that maintains 64 sectors per track, Windows always creates the partition starting at the sixty-forth sector, therefore misaligning it with the underlying physical disk. To be certain of disk alignment, use Diskpart.exe, a disk partition tool. Diskpart.exe is a utility provided by Microsoft in the Windows Server 2003 Service Pack 1 Support Tools that can explicitly set the starting offset in the master boot record (MBR). By setting the starting offset, you can track alignment and improve disk performance. Exchange Server 2003 writes data in multiples of 4 KB I/O operations (4 KB for the databases and up to 32 KB for streaming files). Therefore, make sure that the starting offset is a multiple of 4 KB. Failure to do so may cause a single I/O operation spanning two tracks, causing performance degradation."

Thanks
RWS

ps: for all readers, I never suggested using Raid 5 ... I already own a 6 disk server as discribed in my first post and just want to know the best way to use it

 

[xmsre]

lebisol: back in the exchange 5.5 days, when read/write ratios were 8:1, RAID 5 was fine. In a modern Exchange 2003 environment, the read write ration is 3:1 (uncached clients) or 2:1 (outlook clients in cached mode). The RAID 5 write penalty of 4 is too high in either case.

rws70: The dell install program is a culprit, as it the latest version of IBM ServerGuide. There are two methods to check your allocation unit size (also referred to as cluster size).

1. From the command line run chkdsk with no options. This runs chkdsk in read only mode. Toward the bottom of the output, you'll see a line like "512 bytes in each allocation unit."

2. In the disk defragmentation aplet in the MMC, select the disk and click ?analyze". After the disk is analyzed, click the "veiw report" button. There is no need to actually defragment the disk. In the report, you'll see a line like "cluster size = 512 bytes"

When an application sends an IO request larger than the allocation unit size to the filesystem, NTFS splits the request into multiple IO request of the allocation unit size. You can see this with the perfmon physical disk counter split IOs/sec. If the allocation unit size is 512 bytes, and the application sends a 4K request, NTFS splits it into 8 512byte IOs and sends 8 IOs to the disk.

The same vendors that kill performance with a non-optimal allocation unit size tend to ship with all the disks in one big RAID 5 array as well. With it's write penalty of 4, RAID 5 is suited for applications that read a lot and don't write much. The OS pagefile and system temp directory (both located on the boot volume by default) are just the opposite; they write as much or more than they read. I'm convinced that these vendors intentionally ship the worst possible configuration in terms of disk performance in order to sell unnecessary spindles. One of my customers, a public utility in Arizona, did a study on Dell 28XX as shipped vs. configuring the drives RAID 10 and manually creating the system partition with windows setup; the result was a 4X performance improvement.

This has nothing to do with disk alignment. Disk alignment has been around for a long time. In strict terms it applies only to disks wth a fixed geometry, which have not existed since the days of 9GB SCSI drives. These days, a disk has a variable number of sectors per track ranging between 128 and 250ish. the geometry you see from bios is a fiction created by the disk firmware. Presently, aligning disk tracks really has a lot more to do with aligning cache slots than anything else, and is highly dependent on which specific disk and controller you are using. You don't get the kind of mileage performance you used to three or four years ago due to the advances in disk technology.

 

[rws70]

Follow up:

First I would like to thank xmsre for his very informative posts.

Raid Issue: I have a 6 disk server and I did extensive testing with the exchange Jetstress tool:

http://go.microsoft.com/fwlink/?LinkId=27883

Based on the results I went for this 6 disk setup:

Raid 1 (2 - 146G disks)
C: OS partition and exchange program 92G (cluster size = 512 bytes)
E: Swap partition 9G (cluster size = 4K)
F: Logs partition 35G (cluster size = 4K)

Raid 10 (4 - 76G disks)
G: Info store partition 136G (cluster size = 4K)

(The performance on the logs did not change much putting them on their own raid 1 pair, however the database performance was much better on raid 10. So I put logs on the same raid 1 pair as the OS so I could get the store on raid 10)

I used the Dell server install cd because it loads all the drivers for you, thus my OS partition has a cluster size of 512 bytes. But I created the remaining partitions in windows and thus got the cluster size of 4K which is good for exchange (see posts above).

Hope this helps anyone else with a similar 6 disk server for a <100 user company.

 

[Zelandakh]

Jetstress is all well and good but on a 6 disk system, I'd still have gone for 3 x RAID1 volumes.

2 x 76GB (thought they'd have been 73!) for OS and pagefile
2 x 76GB for logs
2 x 146GB for information store.

You'd have had bigger volumes and greater Exchange performance.

 

[rws70]

Yes they are 73G ... good catch.

I originally was going to go with 3 x Raid1 to get the logs by themselves mainly. But in testing, the logs performance did not change much in each configuration, but the info store was significantly faster on the Raid10.

Thanks for your comments.

 

[Zelandakh]

Good feedback - thanks for the note.

 

[xmsre]

Thanks for the feedback.

John