|
|
|
date: Tue, 10 Jan 2006 13:22:02 -0800,
group: microsoft.public.exchange.design
back
Re: Hardware Recommendations
Another one of my pet peeves. Similar to the old, 'the arrow will never
reach the target' problem. While your number are sound, any decent RAID
controller has on-board cache. writes are cached and delayed until read
are satisfied.
So while the theorical problem points to fsaster writes with mirror, if
you use the same drives and same RAID controler to do R1 or R5, you
should get similar results. Also, 150 users is not much. I run 100 users
on a single 5 drive partitioned as 2 logical drives (OS, Stores and
logs). The performance is more than adequate.
"John Fullbright" wrote in
news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
> I wouldn't do that, "three disks in RAID 5 for the Database .STM and
> .EDB Files it would be 7 Disks.", that is.
>
> Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
> assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
>
> Where P is the performance of a single spindle and N' is the number of
> data spindles in the set:
>
> RAID 5 Read performance = P*N' = 85*2 = 170
> RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
>
> And we apply the read/write raitio which is required when read and
> write performance are asymmetrical. With a 3:1 ratio we are 75% read
> and 25% write:
>
> Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
>
> If you were to buy bigger drives and go with just a mirror:
>
> RAID 1 Read performance = P*N = 170
> RAID 1 Write performance = P*N/2 = 85
>
> Again we apply the read/write ratio:
>
> 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
>
> My recommendation - Three mirrors for a total of 6 spindles. This is
> about 8% better overall performance on the database array with 1 fewer
> spindles. The write performance is a 100% improvement.
>
> Do the math.
>
> John
>
>
>
>
>
> "Jonathan Norris" wrote in
> message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
>> Actually if i were to use two Disks for the OS (Raid 1) and two Disks
>> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
>> Database .STM and
>> .EDB Files it would be 7 Disks.
>>
>> Also keep in mind by having everything on the same spindles you will
>> have an
>> I/O bottleneck.
>>
>> OS/Page File is Read and Write intensive
>> Transaction Logs are write intensive during normal operation and
>> during recovery its read intensive.
>> DB File .stm and .edb are both Read and Write Intensive.
>>
>>
>> --
>> Jonathan
>> No Warrenties Implied, Did you do a FULL backup today??????
>>
>>
>>
>>
>> "Ryan Burrus" wrote:
>>
>>> I am trying to find some information as to the recommended hardware
>>> configuration for Exchange Server 2003. We only have about 150
>>> users and all
>>> access it using Outlook 2003. I have read a few things from
>>> Microsoft that
>>> says the Exchange data files and the page file should be on separate
>>> drives.
>>> Also that the Exchange data files and the Windows system files
>>> should be on
>>> separate drives. Also that the Exchange transaction logs should be
>>> on a separate drive than the Exchange data files. This would equate
>>> to 4 drives
>>> which doesn't sound right to me. Does anyone know what the
>>> recommended scenario would be for an organization our size? Any
>>> advice would be much appreciated. Thanks!!
>
>
>
date: Thu, 12 Jan 2006 12:10:11 -0800
author: Asher_N
Re: Hardware Recommendations
Cache only goes so far. Beyond a base amount, adding more cache is a case
of diminishing returns. If you do rely on cache, it needs to power
protected to prevent corremption in the case of power failure. The main
advantage of cache in RAID 5 is to remove 1 disk access from the 4 required
for each write. This effectively changes the formula for write performance
from P*N'/4 to P*N'/3. Still worse than a mirror.
"Asher_N" wrote in message
news:Xns97499A4FAA3compguy666hotmailcom@207.46.248.16...
> Another one of my pet peeves. Similar to the old, 'the arrow will never
> reach the target' problem. While your number are sound, any decent RAID
> controller has on-board cache. writes are cached and delayed until read
> are satisfied.
>
> So while the theorical problem points to fsaster writes with mirror, if
> you use the same drives and same RAID controler to do R1 or R5, you
> should get similar results. Also, 150 users is not much. I run 100 users
> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
> logs). The performance is more than adequate.
>
> "John Fullbright" wrote in
> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
>
>> I wouldn't do that, "three disks in RAID 5 for the Database .STM and
>> .EDB Files it would be 7 Disks.", that is.
>>
>> Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
>> assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
>>
>> Where P is the performance of a single spindle and N' is the number of
>> data spindles in the set:
>>
>> RAID 5 Read performance = P*N' = 85*2 = 170
>> RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
>>
>> And we apply the read/write raitio which is required when read and
>> write performance are asymmetrical. With a 3:1 ratio we are 75% read
>> and 25% write:
>>
>> Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
>>
>> If you were to buy bigger drives and go with just a mirror:
>>
>> RAID 1 Read performance = P*N = 170
>> RAID 1 Write performance = P*N/2 = 85
>>
>> Again we apply the read/write ratio:
>>
>> 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
>>
>> My recommendation - Three mirrors for a total of 6 spindles. This is
>> about 8% better overall performance on the database array with 1 fewer
>> spindles. The write performance is a 100% improvement.
>>
>> Do the math.
>>
>> John
>>
>>
>>
>>
>>
>> "Jonathan Norris" wrote in
>> message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
>>> Actually if i were to use two Disks for the OS (Raid 1) and two Disks
>>> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
>>> Database .STM and
>>> .EDB Files it would be 7 Disks.
>>>
>>> Also keep in mind by having everything on the same spindles you will
>>> have an
>>> I/O bottleneck.
>>>
>>> OS/Page File is Read and Write intensive
>>> Transaction Logs are write intensive during normal operation and
>>> during recovery its read intensive.
>>> DB File .stm and .edb are both Read and Write Intensive.
>>>
>>>
>>> --
>>> Jonathan
>>> No Warrenties Implied, Did you do a FULL backup today??????
>>>
>>>
>>>
>>>
>>> "Ryan Burrus" wrote:
>>>
>>>> I am trying to find some information as to the recommended hardware
>>>> configuration for Exchange Server 2003. We only have about 150
>>>> users and all
>>>> access it using Outlook 2003. I have read a few things from
>>>> Microsoft that
>>>> says the Exchange data files and the page file should be on separate
>>>> drives.
>>>> Also that the Exchange data files and the Windows system files
>>>> should be on
>>>> separate drives. Also that the Exchange transaction logs should be
>>>> on a separate drive than the Exchange data files. This would equate
>>>> to 4 drives
>>>> which doesn't sound right to me. Does anyone know what the
>>>> recommended scenario would be for an organization our size? Any
>>>> advice would be much appreciated. Thanks!!
>>
>>
>>
>
date: Thu, 12 Jan 2006 16:13:48 -0500
author: John Fullbright
Re: Hardware Recommendations
I am using a calculator developed by Exchange Rangers and have used it
without any issues (both theoretically and in the real world) It uses the
same calculations as John Fullbright had illustrated with more considerations.
I ran the information you provided, if you can provide your average mailbox
size, User Concurrency, Disk Size, Disk RPMs, I will rerun it and give you
the projections.
Interesting enough it doesn't even provide recommendations for RAID1 on the
DB spindles.
Agreed Raid 1 does provide more performance, but I would expect to be given
my pink slip if I recommended it as a customer solution for a Database due to
the lack of redundancy and high risk we would assume.
Unless your users have a really high IOP Profile or really huge mailboxes I
wouldn't expect RAID 5 to be a bottleneck (unless you have your OS and
Transaction Logs sitting on the same Disks/spindles). Which isn't Best
Practice or Recommended by anyone I know (With all due respect).
You may also consider running Microsoft Best Practice Analyser to see if it
gives you any recommendations. Perhaps you have write cache enabled on the
drives?
I would also recommend you run performance monitor to see what the
bottleneck is before you spend money.
Here is a link you may want to check out for Troubleshooting Performance
with Exchange.
http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/e2k3perf.mspx
--
Jonathan
No Warrenties Implied, Did you do a FULL backup today??????
"Asher_N" wrote:
> Another one of my pet peeves. Similar to the old, 'the arrow will never
> reach the target' problem. While your number are sound, any decent RAID
> controller has on-board cache. writes are cached and delayed until read
> are satisfied.
>
> So while the theorical problem points to fsaster writes with mirror, if
> you use the same drives and same RAID controler to do R1 or R5, you
> should get similar results. Also, 150 users is not much. I run 100 users
> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
> logs). The performance is more than adequate.
>
> "John Fullbright" wrote in
> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
>
> > I wouldn't do that, "three disks in RAID 5 for the Database .STM and
> > .EDB Files it would be 7 Disks.", that is.
> >
> > Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
> > assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
> >
> > Where P is the performance of a single spindle and N' is the number of
> > data spindles in the set:
> >
> > RAID 5 Read performance = P*N' = 85*2 = 170
> > RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
> >
> > And we apply the read/write raitio which is required when read and
> > write performance are asymmetrical. With a 3:1 ratio we are 75% read
> > and 25% write:
> >
> > Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
> >
> > If you were to buy bigger drives and go with just a mirror:
> >
> > RAID 1 Read performance = P*N = 170
> > RAID 1 Write performance = P*N/2 = 85
> >
> > Again we apply the read/write ratio:
> >
> > 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
> >
> > My recommendation - Three mirrors for a total of 6 spindles. This is
> > about 8% better overall performance on the database array with 1 fewer
> > spindles. The write performance is a 100% improvement.
> >
> > Do the math.
> >
> > John
> >
> >
> >
> >
> >
> > "Jonathan Norris" wrote in
> > message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
> >> Actually if i were to use two Disks for the OS (Raid 1) and two Disks
> >> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
> >> Database .STM and
> >> .EDB Files it would be 7 Disks.
> >>
> >> Also keep in mind by having everything on the same spindles you will
> >> have an
> >> I/O bottleneck.
> >>
> >> OS/Page File is Read and Write intensive
> >> Transaction Logs are write intensive during normal operation and
> >> during recovery its read intensive.
> >> DB File .stm and .edb are both Read and Write Intensive.
> >>
> >>
> >> --
> >> Jonathan
> >> No Warrenties Implied, Did you do a FULL backup today??????
> >>
> >>
> >>
> >>
> >> "Ryan Burrus" wrote:
> >>
> >>> I am trying to find some information as to the recommended hardware
> >>> configuration for Exchange Server 2003. We only have about 150
> >>> users and all
> >>> access it using Outlook 2003. I have read a few things from
> >>> Microsoft that
> >>> says the Exchange data files and the page file should be on separate
> >>> drives.
> >>> Also that the Exchange data files and the Windows system files
> >>> should be on
> >>> separate drives. Also that the Exchange transaction logs should be
> >>> on a separate drive than the Exchange data files. This would equate
> >>> to 4 drives
> >>> which doesn't sound right to me. Does anyone know what the
> >>> recommended scenario would be for an organization our size? Any
> >>> advice would be much appreciated. Thanks!!
> >
> >
> >
>
>
date: Thu, 12 Jan 2006 13:20:02 -0800
author: Jonathan Norris
Re: Hardware Recommendations
I don't understand what you're talking about when you say " but I would
expect to be given my pink slip if I recommended it as a customer solution
for a Database due to
the lack of redundancy and high risk we would assume."
RAID 1 is 100% redundant, it's mirroring..
The Microsoft Whitepaper "Disk Subsystem Performance Analysis for Windows"
gives a set of criteria where it is appropriate to use RAID 5.
http://download.microsoft.com/download/e/b/a/eba1050f-a31d-436b-9281-92cdfeae4b45/subsys_perf.doc#_Toc72126973
Quoting from the paper:
"RAID-5 might be the right choice for a write-heavy workload if the
following is true:
The workload consists of large requests in comparison to the array's
stripe unit size (that is, they constitute several complete stripes of data
on average), or the controller has a write-back cache, allowing writes to be
delayed and coalesced into full stripes, and
The workload is largely sequential (in terms of LBNs), and
The array controller can detect and optimize for full stripe
writes, and
The number of spindles is reasonable, as described in "Rules of
Thumb" later in this paper, and
Cost is a serious issue"
Exchange databases are a random workload that uses a 4K (small requests) IO
size. A three drive RAID 5 set uses more spindes at higher cost and
provides lower performance than a mirror. I don't see where any of the
criteria are met for a RAID 5 recommendation.
The same paper discusses the characteristics of RAID 1/0+1/10:
" Performance:
For write requests, both parts of a mirror must be updated, reducing
the available write throughput by >50% for 2-way mirrors, >66% for 3-way
mirrors, and so on. On the other hand, read performance improves as
the number of mirrors in each set increases.
Reliability:
For an M-way mirrored set, all M disks must be lost before data is lost. In
the common case of a simple 2-way mirror, MTBF is now increased because it
is the longer of two semi-independent disk lifetimes. In most cases, the
mirror can be restored quickly (for example, by replacing the disk by hand
or using a hot spare), further reducing the window of vulnerability. For
non-trivial mirrored striped arrays, most multi-disk failures are
survivable.
Availability:
If connectivity is lost to one disk, or even M-1 disks of every M-mirrored
set in a striped array, work can still proceed, albeit at a different level
of throughput; failure of a mirror spindle decreases read performance and
increases write performance for that set. When disconnected disks become
reachable again, they can be synchronized back to the corresponding active
disks, thereby restoring the original array characteristics."
Both a RAID 5 set with 3 disks and a Mirror can withstand the failure of a
single drive. I'm not seeing a redundancy problem that would result in a
pink slip here. I do see a potential perfrormance problem with RAID 5 that
could however result in that pink slip.
"Jonathan Norris" wrote in
message news:896AC5A6-9D41-412C-8498-265B7D2A11AB@microsoft.com...
>I am using a calculator developed by Exchange Rangers and have used it
> without any issues (both theoretically and in the real world) It uses the
> same calculations as John Fullbright had illustrated with more
> considerations.
>
> I ran the information you provided, if you can provide your average
> mailbox
> size, User Concurrency, Disk Size, Disk RPMs, I will rerun it and give you
> the projections.
>
> Interesting enough it doesn't even provide recommendations for RAID1 on
> the
> DB spindles.
>
> Agreed Raid 1 does provide more performance, but I would expect to be
> given
> my pink slip if I recommended it as a customer solution for a Database due
> to
> the lack of redundancy and high risk we would assume.
>
> Unless your users have a really high IOP Profile or really huge mailboxes
> I
> wouldn't expect RAID 5 to be a bottleneck (unless you have your OS and
> Transaction Logs sitting on the same Disks/spindles). Which isn't Best
> Practice or Recommended by anyone I know (With all due respect).
>
> You may also consider running Microsoft Best Practice Analyser to see if
> it
> gives you any recommendations. Perhaps you have write cache enabled on
> the
> drives?
>
> I would also recommend you run performance monitor to see what the
> bottleneck is before you spend money.
>
> Here is a link you may want to check out for Troubleshooting Performance
> with Exchange.
>
> http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/e2k3perf.mspx
>
> --
> Jonathan
> No Warrenties Implied, Did you do a FULL backup today??????
>
>
>
>
> "Asher_N" wrote:
>
>> Another one of my pet peeves. Similar to the old, 'the arrow will never
>> reach the target' problem. While your number are sound, any decent RAID
>> controller has on-board cache. writes are cached and delayed until read
>> are satisfied.
>>
>> So while the theorical problem points to fsaster writes with mirror, if
>> you use the same drives and same RAID controler to do R1 or R5, you
>> should get similar results. Also, 150 users is not much. I run 100 users
>> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
>> logs). The performance is more than adequate.
>>
>> "John Fullbright" wrote in
>> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
>>
>> > I wouldn't do that, "three disks in RAID 5 for the Database .STM and
>> > .EDB Files it would be 7 Disks.", that is.
>> >
>> > Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
>> > assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
>> >
>> > Where P is the performance of a single spindle and N' is the number of
>> > data spindles in the set:
>> >
>> > RAID 5 Read performance = P*N' = 85*2 = 170
>> > RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
>> >
>> > And we apply the read/write raitio which is required when read and
>> > write performance are asymmetrical. With a 3:1 ratio we are 75% read
>> > and 25% write:
>> >
>> > Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
>> >
>> > If you were to buy bigger drives and go with just a mirror:
>> >
>> > RAID 1 Read performance = P*N = 170
>> > RAID 1 Write performance = P*N/2 = 85
>> >
>> > Again we apply the read/write ratio:
>> >
>> > 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
>> >
>> > My recommendation - Three mirrors for a total of 6 spindles. This is
>> > about 8% better overall performance on the database array with 1 fewer
>> > spindles. The write performance is a 100% improvement.
>> >
>> > Do the math.
>> >
>> > John
>> >
>> >
>> >
>> >
>> >
>> > "Jonathan Norris" wrote in
>> > message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
>> >> Actually if i were to use two Disks for the OS (Raid 1) and two Disks
>> >> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
>> >> Database .STM and
>> >> .EDB Files it would be 7 Disks.
>> >>
>> >> Also keep in mind by having everything on the same spindles you will
>> >> have an
>> >> I/O bottleneck.
>> >>
>> >> OS/Page File is Read and Write intensive
>> >> Transaction Logs are write intensive during normal operation and
>> >> during recovery its read intensive.
>> >> DB File .stm and .edb are both Read and Write Intensive.
>> >>
>> >>
>> >> --
>> >> Jonathan
>> >> No Warrenties Implied, Did you do a FULL backup today??????
>> >>
>> >>
>> >>
>> >>
>> >> "Ryan Burrus" wrote:
>> >>
>> >>> I am trying to find some information as to the recommended hardware
>> >>> configuration for Exchange Server 2003. We only have about 150
>> >>> users and all
>> >>> access it using Outlook 2003. I have read a few things from
>> >>> Microsoft that
>> >>> says the Exchange data files and the page file should be on separate
>> >>> drives.
>> >>> Also that the Exchange data files and the Windows system files
>> >>> should be on
>> >>> separate drives. Also that the Exchange transaction logs should be
>> >>> on a separate drive than the Exchange data files. This would equate
>> >>> to 4 drives
>> >>> which doesn't sound right to me. Does anyone know what the
>> >>> recommended scenario would be for an organization our size? Any
>> >>> advice would be much appreciated. Thanks!!
>> >
>> >
>> >
>>
>>
date: Thu, 12 Jan 2006 16:47:22 -0500
author: John Fullbright
Re: Hardware Recommendations
I got this from
http://www.microsoft.com/technet/prodtechnol/exchange/guides/StoragePerformance/1c471676-2312-4ffe-adf2-15a9cfd529c4.mspx
The RAID solution you use should be based on the cost and performance
trade-offs that are appropriate for your company. As a result, in many cases,
more than one type of RAID solution may be recommended for a particular data
storage requirement. General recommendations are as follows:
• Use Raid-1+0 for the transaction log file volumes, database volumes, and
SMTP queues
• Use Raid-1 for the transaction log file volumes, SMTP queues, and MTA
queues.
• In general, Raid-5 does not provide the best trade-off between
reliability/availability and performance.
• Raid-0 is never recommended.
The only concern I have with running RAID 1 is if you do loose a disk you
will have to down the server to rebuild it unless the RAID controller
supports hot swap, then again if you loose the disk in the Transaction Logs
or OS your just as dead in the water.
Interesting enough they don't come out and say RAID-1 is suitable for the
DBs in any part of the article.
Perhaps MS should write another article for the SMB segment. Since Raid 0+1
isn't a cost effective solution for 150 users its open for debate.
I am not disagreeing with you that RAID 1 provides more performance than
RAID 5.
However running the numbers... RAID 5 will support the number of users and
the IOPS required to support this many users. But RAID 1 would use less
disks and provide more performance.
So your solution would be better in this case. I spend so little time in
the SMB segment that I never implement this scenerio. I stand corrected:)
Here is another article for the person who posted the question,
http://www.microsoft.com/technet/prodtechnol/exchange/guides/E2k3HighAvGuide/c1804c14-c099-4207-b6b9-de5bda972b76.mspx
--
Jonathan
No Warrenties Implied, Did you do a FULL backup today??????
"John Fullbright" wrote:
> I don't understand what you're talking about when you say " but I would
> expect to be given my pink slip if I recommended it as a customer solution
> for a Database due to
> the lack of redundancy and high risk we would assume."
>
> RAID 1 is 100% redundant, it's mirroring..
>
> The Microsoft Whitepaper "Disk Subsystem Performance Analysis for Windows"
> gives a set of criteria where it is appropriate to use RAID 5.
> http://download.microsoft.com/download/e/b/a/eba1050f-a31d-436b-9281-92cdfeae4b45/subsys_perf.doc#_Toc72126973
> Quoting from the paper:
>
> "RAID-5 might be the right choice for a write-heavy workload if the
> following is true:
>
> · The workload consists of large requests in comparison to the array's
> stripe unit size (that is, they constitute several complete stripes of data
> on average), or the controller has a write-back cache, allowing writes to be
> delayed and coalesced into full stripes, and
>
> · The workload is largely sequential (in terms of LBNs), and
>
> · The array controller can detect and optimize for full stripe
> writes, and
>
> · The number of spindles is reasonable, as described in "Rules of
> Thumb" later in this paper, and
>
> · Cost is a serious issue"
>
>
> Exchange databases are a random workload that uses a 4K (small requests) IO
> size. A three drive RAID 5 set uses more spindes at higher cost and
> provides lower performance than a mirror. I don't see where any of the
> criteria are met for a RAID 5 recommendation.
>
> The same paper discusses the characteristics of RAID 1/0+1/10:
>
> " · Performance:
> For write requests, both parts of a mirror must be updated, reducing
> the available write throughput by >50% for 2-way mirrors, >66% for 3-way
> mirrors, and so on. On the other hand, read performance improves as
> the number of mirrors in each set increases.
> · Reliability:
> For an M-way mirrored set, all M disks must be lost before data is lost. In
> the common case of a simple 2-way mirror, MTBF is now increased because it
> is the longer of two semi-independent disk lifetimes. In most cases, the
> mirror can be restored quickly (for example, by replacing the disk by hand
> or using a hot spare), further reducing the window of vulnerability. For
> non-trivial mirrored striped arrays, most multi-disk failures are
> survivable.
>
> · Availability:
> If connectivity is lost to one disk, or even M-1 disks of every M-mirrored
> set in a striped array, work can still proceed, albeit at a different level
> of throughput; failure of a mirror spindle decreases read performance and
> increases write performance for that set. When disconnected disks become
> reachable again, they can be synchronized back to the corresponding active
> disks, thereby restoring the original array characteristics."
>
>
> Both a RAID 5 set with 3 disks and a Mirror can withstand the failure of a
> single drive. I'm not seeing a redundancy problem that would result in a
> pink slip here. I do see a potential perfrormance problem with RAID 5 that
> could however result in that pink slip.
>
>
>
>
>
> "Jonathan Norris" wrote in
> message news:896AC5A6-9D41-412C-8498-265B7D2A11AB@microsoft.com...
> >I am using a calculator developed by Exchange Rangers and have used it
> > without any issues (both theoretically and in the real world) It uses the
> > same calculations as John Fullbright had illustrated with more
> > considerations.
> >
> > I ran the information you provided, if you can provide your average
> > mailbox
> > size, User Concurrency, Disk Size, Disk RPMs, I will rerun it and give you
> > the projections.
> >
> > Interesting enough it doesn't even provide recommendations for RAID1 on
> > the
> > DB spindles.
> >
> > Agreed Raid 1 does provide more performance, but I would expect to be
> > given
> > my pink slip if I recommended it as a customer solution for a Database due
> > to
> > the lack of redundancy and high risk we would assume.
> >
> > Unless your users have a really high IOP Profile or really huge mailboxes
> > I
> > wouldn't expect RAID 5 to be a bottleneck (unless you have your OS and
> > Transaction Logs sitting on the same Disks/spindles). Which isn't Best
> > Practice or Recommended by anyone I know (With all due respect).
> >
> > You may also consider running Microsoft Best Practice Analyser to see if
> > it
> > gives you any recommendations. Perhaps you have write cache enabled on
> > the
> > drives?
> >
> > I would also recommend you run performance monitor to see what the
> > bottleneck is before you spend money.
> >
> > Here is a link you may want to check out for Troubleshooting Performance
> > with Exchange.
> >
> > http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/e2k3perf.mspx
> >
> > --
> > Jonathan
> > No Warrenties Implied, Did you do a FULL backup today??????
> >
> >
> >
> >
> > "Asher_N" wrote:
> >
> >> Another one of my pet peeves. Similar to the old, 'the arrow will never
> >> reach the target' problem. While your number are sound, any decent RAID
> >> controller has on-board cache. writes are cached and delayed until read
> >> are satisfied.
> >>
> >> So while the theorical problem points to fsaster writes with mirror, if
> >> you use the same drives and same RAID controler to do R1 or R5, you
> >> should get similar results. Also, 150 users is not much. I run 100 users
> >> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
> >> logs). The performance is more than adequate.
> >>
> >> "John Fullbright" wrote in
> >> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
> >>
> >> > I wouldn't do that, "three disks in RAID 5 for the Database .STM and
> >> > .EDB Files it would be 7 Disks.", that is.
> >> >
> >> > Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
> >> > assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
> >> >
> >> > Where P is the performance of a single spindle and N' is the number of
> >> > data spindles in the set:
> >> >
> >> > RAID 5 Read performance = P*N' = 85*2 = 170
> >> > RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
> >> >
> >> > And we apply the read/write raitio which is required when read and
> >> > write performance are asymmetrical. With a 3:1 ratio we are 75% read
> >> > and 25% write:
> >> >
> >> > Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
> >> >
> >> > If you were to buy bigger drives and go with just a mirror:
> >> >
> >> > RAID 1 Read performance = P*N = 170
> >> > RAID 1 Write performance = P*N/2 = 85
> >> >
> >> > Again we apply the read/write ratio:
> >> >
> >> > 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
> >> >
> >> > My recommendation - Three mirrors for a total of 6 spindles. This is
> >> > about 8% better overall performance on the database array with 1 fewer
> >> > spindles. The write performance is a 100% improvement.
> >> >
> >> > Do the math.
> >> >
> >> > John
> >> >
> >> >
> >> >
> >> >
> >> >
> >> > "Jonathan Norris" wrote in
> >> > message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
> >> >> Actually if i were to use two Disks for the OS (Raid 1) and two Disks
> >> >> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
> >> >> Database .STM and
> >> >> .EDB Files it would be 7 Disks.
> >> >>
> >> >> Also keep in mind by having everything on the same spindles you will
> >> >> have an
> >> >> I/O bottleneck.
> >> >>
> >> >> OS/Page File is Read and Write intensive
> >> >> Transaction Logs are write intensive during normal operation and
> >> >> during recovery its read intensive.
> >> >> DB File .stm and .edb are both Read and Write Intensive.
> >> >>
> >> >>
> >> >> --
> >> >> Jonathan
> >> >> No Warrenties Implied, Did you do a FULL backup today??????
> >> >>
> >> >>
> >> >>
> >> >>
> >> >> "Ryan Burrus" wrote:
> >> >>
> >> >>> I am trying to find some information as to the recommended hardware
> >> >>> configuration for Exchange Server 2003. We only have about 150
> >> >>> users and all
> >> >>> access it using Outlook 2003. I have read a few things from
> >> >>> Microsoft that
> >> >>> says the Exchange data files and the page file should be on separate
> >> >>> drives.
> >> >>> Also that the Exchange data files and the Windows system files
> >> >>> should be on
> >> >>> separate drives. Also that the Exchange transaction logs should be
> >> >>> on a separate drive than the Exchange data files. This would equate
> >> >>> to 4 drives
> >> >>> which doesn't sound right to me. Does anyone know what the
> >> >>> recommended scenario would be for an organization our size? Any
> >> >>> advice would be much appreciated. Thanks!!
> >> >
> >> >
> >> >
> >>
> >>
>
>
>
date: Fri, 13 Jan 2006 09:07:02 -0800
author: Jonathan Norris
Re: Hardware Recommendations
"Interesting enough they don't come out and say RAID-1 is suitable for the
DBs in any part of the article.
Perhaps MS should write another article for the SMB segment. Since Raid 0+1
isn't a cost effective solution for 150 users its open for debate.
I am not disagreeing with you that RAID 1 provides more performance than
RAID 5."
Jonathan, I see that you came to the conclusion that RAID 1 was pretty good
in this case. Since you're on the learning track, can I suggest that you
rethink the RAID concepts as you've come to know them? Also, rethink the
way that controllers deal with RAID. Cache is fine but keep in mind that
Exchange is cache-unfriendly.
Also, when you read those documents that folks write, keep in mind they
write them for the general audience. Although I've personally seen some of
the documents for the SMB segment (and they're pretty good if I say so
myself), don't expect Microsoft to publish papers that say that RAID 1 is
recommended for Exchange. Will it work? Yep. Is it as good as RAID 5? For
some types of operations, yes. For others, no. If you have read heavy
applications, you'll get better performance from a RAID 5 than you will from
RAID 1. Write performance would be less. Both would have the same type of
fault tolerance. I throw this out there hopefully to be helpful. Besides,
you wouldn't have to count on tools from Rangers that way ;)
Johnathan (?), I'm surprised to see in your calculation a number of 2 for
the number of spindles in the data set. "RAID 5 Read performance = P*N' =
85*2 = 170" Can you expand on that? I see two, but it's a raid5 set
meaning you have three possible targets. Are you saying that only two are
valid for both read and writes? Traditionally, I would have expected 3 for
the write and 2 for the read to account for the parity penalty. Curious. I
realize you're the guy to know, but I'm interested in knowing why that
number is used vs. another.
Cheers,
Al
"Jonathan Norris" wrote in
message news:8EA67CFF-43E2-41FB-B0C1-76BF87504AD0@microsoft.com...
>I got this from
> http://www.microsoft.com/technet/prodtechnol/exchange/guides/StoragePerformance/1c471676-2312-4ffe-adf2-15a9cfd529c4.mspx
>
> The RAID solution you use should be based on the cost and performance
> trade-offs that are appropriate for your company. As a result, in many
> cases,
> more than one type of RAID solution may be recommended for a particular
> data
> storage requirement. General recommendations are as follows:
>
> . Use Raid-1+0 for the transaction log file volumes, database volumes, and
> SMTP queues
>
> . Use Raid-1 for the transaction log file volumes, SMTP queues, and MTA
> queues.
>
>
> . In general, Raid-5 does not provide the best trade-off between
> reliability/availability and performance.
>
> . Raid-0 is never recommended.
>
> The only concern I have with running RAID 1 is if you do loose a disk you
> will have to down the server to rebuild it unless the RAID controller
> supports hot swap, then again if you loose the disk in the Transaction
> Logs
> or OS your just as dead in the water.
>
> Interesting enough they don't come out and say RAID-1 is suitable for the
> DBs in any part of the article.
>
> Perhaps MS should write another article for the SMB segment. Since Raid
> 0+1
> isn't a cost effective solution for 150 users its open for debate.
>
> I am not disagreeing with you that RAID 1 provides more performance than
> RAID 5.
>
> However running the numbers... RAID 5 will support the number of users and
> the IOPS required to support this many users. But RAID 1 would use less
> disks and provide more performance.
>
> So your solution would be better in this case. I spend so little time in
> the SMB segment that I never implement this scenerio. I stand corrected:)
>
> Here is another article for the person who posted the question,
>
> http://www.microsoft.com/technet/prodtechnol/exchange/guides/E2k3HighAvGuide/c1804c14-c099-4207-b6b9-de5bda972b76.mspx
>
>
>
>
>
>
>
>
>
>
>
> --
> Jonathan
> No Warrenties Implied, Did you do a FULL backup today??????
>
>
>
>
> "John Fullbright" wrote:
>
>> I don't understand what you're talking about when you say " but I would
>> expect to be given my pink slip if I recommended it as a customer
>> solution
>> for a Database due to
>> the lack of redundancy and high risk we would assume."
>>
>> RAID 1 is 100% redundant, it's mirroring..
>>
>> The Microsoft Whitepaper "Disk Subsystem Performance Analysis for
>> Windows"
>> gives a set of criteria where it is appropriate to use RAID 5.
>> http://download.microsoft.com/download/e/b/a/eba1050f-a31d-436b-9281-92cdfeae4b45/subsys_perf.doc#_Toc72126973
>> Quoting from the paper:
>>
>> "RAID-5 might be the right choice for a write-heavy workload if the
>> following is true:
>>
>> The workload consists of large requests in comparison to the
>> array's
>> stripe unit size (that is, they constitute several complete stripes of
>> data
>> on average), or the controller has a write-back cache, allowing writes to
>> be
>> delayed and coalesced into full stripes, and
>>
>> The workload is largely sequential (in terms of LBNs), and
>>
>> The array controller can detect and optimize for full stripe
>> writes, and
>>
>> The number of spindles is reasonable, as described in "Rules of
>> Thumb" later in this paper, and
>>
>> Cost is a serious issue"
>>
>>
>> Exchange databases are a random workload that uses a 4K (small requests)
>> IO
>> size. A three drive RAID 5 set uses more spindes at higher cost and
>> provides lower performance than a mirror. I don't see where any of the
>> criteria are met for a RAID 5 recommendation.
>>
>> The same paper discusses the characteristics of RAID 1/0+1/10:
>>
>> " Performance:
>> For write requests, both parts of a mirror must be updated,
>> reducing
>> the available write throughput by >50% for 2-way mirrors, >66% for 3-way
>> mirrors, and so on. On the other hand, read performance improves
>> as
>> the number of mirrors in each set increases.
>> Reliability:
>> For an M-way mirrored set, all M disks must be lost before data is lost.
>> In
>> the common case of a simple 2-way mirror, MTBF is now increased because
>> it
>> is the longer of two semi-independent disk lifetimes. In most cases, the
>> mirror can be restored quickly (for example, by replacing the disk by
>> hand
>> or using a hot spare), further reducing the window of vulnerability. For
>> non-trivial mirrored striped arrays, most multi-disk failures are
>> survivable.
>>
>> Availability:
>> If connectivity is lost to one disk, or even M-1 disks of every
>> M-mirrored
>> set in a striped array, work can still proceed, albeit at a different
>> level
>> of throughput; failure of a mirror spindle decreases read performance and
>> increases write performance for that set. When disconnected disks become
>> reachable again, they can be synchronized back to the corresponding
>> active
>> disks, thereby restoring the original array characteristics."
>>
>>
>> Both a RAID 5 set with 3 disks and a Mirror can withstand the failure of
>> a
>> single drive. I'm not seeing a redundancy problem that would result in a
>> pink slip here. I do see a potential perfrormance problem with RAID 5
>> that
>> could however result in that pink slip.
>>
>>
>>
>>
>>
>> "Jonathan Norris" wrote in
>> message news:896AC5A6-9D41-412C-8498-265B7D2A11AB@microsoft.com...
>> >I am using a calculator developed by Exchange Rangers and have used it
>> > without any issues (both theoretically and in the real world) It uses
>> > the
>> > same calculations as John Fullbright had illustrated with more
>> > considerations.
>> >
>> > I ran the information you provided, if you can provide your average
>> > mailbox
>> > size, User Concurrency, Disk Size, Disk RPMs, I will rerun it and give
>> > you
>> > the projections.
>> >
>> > Interesting enough it doesn't even provide recommendations for RAID1 on
>> > the
>> > DB spindles.
>> >
>> > Agreed Raid 1 does provide more performance, but I would expect to be
>> > given
>> > my pink slip if I recommended it as a customer solution for a Database
>> > due
>> > to
>> > the lack of redundancy and high risk we would assume.
>> >
>> > Unless your users have a really high IOP Profile or really huge
>> > mailboxes
>> > I
>> > wouldn't expect RAID 5 to be a bottleneck (unless you have your OS and
>> > Transaction Logs sitting on the same Disks/spindles). Which isn't Best
>> > Practice or Recommended by anyone I know (With all due respect).
>> >
>> > You may also consider running Microsoft Best Practice Analyser to see
>> > if
>> > it
>> > gives you any recommendations. Perhaps you have write cache enabled on
>> > the
>> > drives?
>> >
>> > I would also recommend you run performance monitor to see what the
>> > bottleneck is before you spend money.
>> >
>> > Here is a link you may want to check out for Troubleshooting
>> > Performance
>> > with Exchange.
>> >
>> > http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/e2k3perf.mspx
>> >
>> > --
>> > Jonathan
>> > No Warrenties Implied, Did you do a FULL backup today??????
>> >
>> >
>> >
>> >
>> > "Asher_N" wrote:
>> >
>> >> Another one of my pet peeves. Similar to the old, 'the arrow will
>> >> never
>> >> reach the target' problem. While your number are sound, any decent
>> >> RAID
>> >> controller has on-board cache. writes are cached and delayed until
>> >> read
>> >> are satisfied.
>> >>
>> >> So while the theorical problem points to fsaster writes with mirror,
>> >> if
>> >> you use the same drives and same RAID controler to do R1 or R5, you
>> >> should get similar results. Also, 150 users is not much. I run 100
>> >> users
>> >> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
>> >> logs). The performance is more than adequate.
>> >>
>> >> "John Fullbright" wrote in
>> >> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
>> >>
>> >> > I wouldn't do that, "three disks in RAID 5 for the Database .STM and
>> >> > .EDB Files it would be 7 Disks.", that is.
>> >> >
>> >> > Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
>> >> > assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
>> >> >
>> >> > Where P is the performance of a single spindle and N' is the number
>> >> > of
>> >> > data spindles in the set:
>> >> >
>> >> > RAID 5 Read performance = P*N' = 85*2 = 170
>> >> > RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
>> >> >
>> >> > And we apply the read/write raitio which is required when read and
>> >> > write performance are asymmetrical. With a 3:1 ratio we are 75%
>> >> > read
>> >> > and 25% write:
>> >> >
>> >> > Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
>> >> >
>> >> > If you were to buy bigger drives and go with just a mirror:
>> >> >
>> >> > RAID 1 Read performance = P*N = 170
>> >> > RAID 1 Write performance = P*N/2 = 85
>> >> >
>> >> > Again we apply the read/write ratio:
>> >> >
>> >> > 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
>> >> >
>> >> > My recommendation - Three mirrors for a total of 6 spindles. This
>> >> > is
>> >> > about 8% better overall performance on the database array with 1
>> >> > fewer
>> >> > spindles. The write performance is a 100% improvement.
>> >> >
>> >> > Do the math.
>> >> >
>> >> > John
>> >> >
>> >> >
>> >> >
>> >> >
>> >> >
>> >> > "Jonathan Norris" wrote
>> >> > in
>> >> > message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
>> >> >> Actually if i were to use two Disks for the OS (Raid 1) and two
>> >> >> Disks
>> >> >> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
>> >> >> Database .STM and
>> >> >> .EDB Files it would be 7 Disks.
>> >> >>
>> >> >> Also keep in mind by having everything on the same spindles you
>> >> >> will
>> >> >> have an
>> >> >> I/O bottleneck.
>> >> >>
>> >> >> OS/Page File is Read and Write intensive
>> >> >> Transaction Logs are write intensive during normal operation and
>> >> >> during recovery its read intensive.
>> >> >> DB File .stm and .edb are both Read and Write Intensive.
>> >> >>
>> >> >>
>> >> >> --
>> >> >> Jonathan
>> >> >> No Warrenties Implied, Did you do a FULL backup today??????
>> >> >>
>> >> >>
>> >> >>
>> >> >>
>> >> >> "Ryan Burrus" wrote:
>> >> >>
>> >> >>> I am trying to find some information as to the recommended
>> >> >>> hardware
>> >> >>> configuration for Exchange Server 2003. We only have about 150
>> >> >>> users and all
>> >> >>> access it using Outlook 2003. I have read a few things from
>> >> >>> Microsoft that
>> >> >>> says the Exchange data files and the page file should be on
>> >> >>> separate
>> >> >>> drives.
>> >> >>> Also that the Exchange data files and the Windows system files
>> >> >>> should be on
>> >> >>> separate drives. Also that the Exchange transaction logs should
>> >> >>> be
>> >> >>> on a separate drive than the Exchange data files. This would
>> >> >>> equate
>> >> >>> to 4 drives
>> >> >>> which doesn't sound right to me. Does anyone know what the
>> >> >>> recommended scenario would be for an organization our size? Any
>> >> >>> advice would be much appreciated. Thanks!!
>> >> >
>> >> >
>> >> >
>> >>
>> >>
>>
>>
>>
date: Wed, 18 Jan 2006 21:38:19 -0500
author: Al Mulnick
Re: Hardware Recommendations
Al Mulnick. Wow, another blast from the past. Still hangn' with Jim at
Bubba's Bait & Tackle?
Anyway:
P = Performance of a single spindle in IOPS
N = Number of spindles in the RAID set.
N' = Number of data spindles in the RAID set (N - parity spindles)
RAID 5 performance calculations:
Write performance = P*N'/4
Read performance = P*N'
(depending on the caching architecture implemented by the controller and the
amount of write cache, in some cases write performance can be expressed
P*N'/3)
RAID 1/10/0+1 performance caluclations:
Write performance = P*N/2
Read performance = P*N
RAID 4/RAID DP (as implemented by Network Appliance) performance
calculations
Write performance = P*N'
Read performance = P*N'
For a 3 drive RAID 5 set, N' would be 2. The set size would be three and
the other spindle is parity.
"Al Mulnick" wrote in message
news:uqljpFKHGHA.1628@TK2MSFTNGP12.phx.gbl...
> "Interesting enough they don't come out and say RAID-1 is suitable for the
> DBs in any part of the article.
>
> Perhaps MS should write another article for the SMB segment. Since Raid
> 0+1
> isn't a cost effective solution for 150 users its open for debate.
>
> I am not disagreeing with you that RAID 1 provides more performance than
> RAID 5."
>
> Jonathan, I see that you came to the conclusion that RAID 1 was pretty
> good in this case. Since you're on the learning track, can I suggest that
> you rethink the RAID concepts as you've come to know them? Also, rethink
> the way that controllers deal with RAID. Cache is fine but keep in mind
> that Exchange is cache-unfriendly.
>
> Also, when you read those documents that folks write, keep in mind they
> write them for the general audience. Although I've personally seen some
> of the documents for the SMB segment (and they're pretty good if I say so
> myself), don't expect Microsoft to publish papers that say that RAID 1 is
> recommended for Exchange. Will it work? Yep. Is it as good as RAID 5?
> For some types of operations, yes. For others, no. If you have read
> heavy applications, you'll get better performance from a RAID 5 than you
> will from RAID 1. Write performance would be less. Both would have the
> same type of fault tolerance. I throw this out there hopefully to be
> helpful. Besides, you wouldn't have to count on tools from Rangers that
> way ;)
>
> Johnathan (?), I'm surprised to see in your calculation a number of 2 for
> the number of spindles in the data set. "RAID 5 Read performance = P*N' =
> 85*2 = 170" Can you expand on that? I see two, but it's a raid5 set
> meaning you have three possible targets. Are you saying that only two are
> valid for both read and writes? Traditionally, I would have expected 3 for
> the write and 2 for the read to account for the parity penalty. Curious. I
> realize you're the guy to know, but I'm interested in knowing why that
> number is used vs. another.
>
> Cheers,
> Al
>
>
> "Jonathan Norris" wrote in
> message news:8EA67CFF-43E2-41FB-B0C1-76BF87504AD0@microsoft.com...
>>I got this from
>> http://www.microsoft.com/technet/prodtechnol/exchange/guides/StoragePerformance/1c471676-2312-4ffe-adf2-15a9cfd529c4.mspx
>>
>> The RAID solution you use should be based on the cost and performance
>> trade-offs that are appropriate for your company. As a result, in many
>> cases,
>> more than one type of RAID solution may be recommended for a particular
>> data
>> storage requirement. General recommendations are as follows:
>>
>> . Use Raid-1+0 for the transaction log file volumes, database volumes,
>> and
>> SMTP queues
>>
>> . Use Raid-1 for the transaction log file volumes, SMTP queues, and MTA
>> queues.
>>
>>
>> . In general, Raid-5 does not provide the best trade-off between
>> reliability/availability and performance.
>>
>> . Raid-0 is never recommended.
>>
>> The only concern I have with running RAID 1 is if you do loose a disk you
>> will have to down the server to rebuild it unless the RAID controller
>> supports hot swap, then again if you loose the disk in the Transaction
>> Logs
>> or OS your just as dead in the water.
>>
>> Interesting enough they don't come out and say RAID-1 is suitable for the
>> DBs in any part of the article.
>>
>> Perhaps MS should write another article for the SMB segment. Since Raid
>> 0+1
>> isn't a cost effective solution for 150 users its open for debate.
>>
>> I am not disagreeing with you that RAID 1 provides more performance than
>> RAID 5.
>>
>> However running the numbers... RAID 5 will support the number of users
>> and
>> the IOPS required to support this many users. But RAID 1 would use less
>> disks and provide more performance.
>>
>> So your solution would be better in this case. I spend so little time in
>> the SMB segment that I never implement this scenerio. I stand
>> corrected:)
>>
>> Here is another article for the person who posted the question,
>>
>> http://www.microsoft.com/technet/prodtechnol/exchange/guides/E2k3HighAvGuide/c1804c14-c099-4207-b6b9-de5bda972b76.mspx
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> --
>> Jonathan
>> No Warrenties Implied, Did you do a FULL backup today??????
>>
>>
>>
>>
>> "John Fullbright" wrote:
>>
>>> I don't understand what you're talking about when you say " but I would
>>> expect to be given my pink slip if I recommended it as a customer
>>> solution
>>> for a Database due to
>>> the lack of redundancy and high risk we would assume."
>>>
>>> RAID 1 is 100% redundant, it's mirroring..
>>>
>>> The Microsoft Whitepaper "Disk Subsystem Performance Analysis for
>>> Windows"
>>> gives a set of criteria where it is appropriate to use RAID 5.
>>> http://download.microsoft.com/download/e/b/a/eba1050f-a31d-436b-9281-92cdfeae4b45/subsys_perf.doc#_Toc72126973
>>> Quoting from the paper:
>>>
>>> "RAID-5 might be the right choice for a write-heavy workload if the
>>> following is true:
>>>
>>> The workload consists of large requests in comparison to the
>>> array's
>>> stripe unit size (that is, they constitute several complete stripes of
>>> data
>>> on average), or the controller has a write-back cache, allowing writes
>>> to be
>>> delayed and coalesced into full stripes, and
>>>
>>> The workload is largely sequential (in terms of LBNs), and
>>>
>>> The array controller can detect and optimize for full stripe
>>> writes, and
>>>
>>> The number of spindles is reasonable, as described in "Rules
>>> of
>>> Thumb" later in this paper, and
>>>
>>> Cost is a serious issue"
>>>
>>>
>>> Exchange databases are a random workload that uses a 4K (small requests)
>>> IO
>>> size. A three drive RAID 5 set uses more spindes at higher cost and
>>> provides lower performance than a mirror. I don't see where any of the
>>> criteria are met for a RAID 5 recommendation.
>>>
>>> The same paper discusses the characteristics of RAID 1/0+1/10:
>>>
>>> " Performance:
>>> For write requests, both parts of a mirror must be updated,
>>> reducing
>>> the available write throughput by >50% for 2-way mirrors, >66% for 3-way
>>> mirrors, and so on. On the other hand, read performance improves as
>>> the number of mirrors in each set increases.
>>> Reliability:
>>> For an M-way mirrored set, all M disks must be lost before data is lost.
>>> In
>>> the common case of a simple 2-way mirror, MTBF is now increased because
>>> it
>>> is the longer of two semi-independent disk lifetimes. In most cases, the
>>> mirror can be restored quickly (for example, by replacing the disk by
>>> hand
>>> or using a hot spare), further reducing the window of vulnerability. For
>>> non-trivial mirrored striped arrays, most multi-disk failures are
>>> survivable.
>>>
>>> Availability:
>>> If connectivity is lost to one disk, or even M-1 disks of every
>>> M-mirrored
>>> set in a striped array, work can still proceed, albeit at a different
>>> level
>>> of throughput; failure of a mirror spindle decreases read performance
>>> and
>>> increases write performance for that set. When disconnected disks become
>>> reachable again, they can be synchronized back to the corresponding
>>> active
>>> disks, thereby restoring the original array characteristics."
>>>
>>>
>>> Both a RAID 5 set with 3 disks and a Mirror can withstand the failure of
>>> a
>>> single drive. I'm not seeing a redundancy problem that would result in
>>> a
>>> pink slip here. I do see a potential perfrormance problem with RAID 5
>>> that
>>> could however result in that pink slip.
>>>
>>>
>>>
>>>
>>>
>>> "Jonathan Norris" wrote in
>>> message news:896AC5A6-9D41-412C-8498-265B7D2A11AB@microsoft.com...
>>> >I am using a calculator developed by Exchange Rangers and have used it
>>> > without any issues (both theoretically and in the real world) It uses
>>> > the
>>> > same calculations as John Fullbright had illustrated with more
>>> > considerations.
>>> >
>>> > I ran the information you provided, if you can provide your average
>>> > mailbox
>>> > size, User Concurrency, Disk Size, Disk RPMs, I will rerun it and give
>>> > you
>>> > the projections.
>>> >
>>> > Interesting enough it doesn't even provide recommendations for RAID1
>>> > on
>>> > the
>>> > DB spindles.
>>> >
>>> > Agreed Raid 1 does provide more performance, but I would expect to be
>>> > given
>>> > my pink slip if I recommended it as a customer solution for a Database
>>> > due
>>> > to
>>> > the lack of redundancy and high risk we would assume.
>>> >
>>> > Unless your users have a really high IOP Profile or really huge
>>> > mailboxes
>>> > I
>>> > wouldn't expect RAID 5 to be a bottleneck (unless you have your OS and
>>> > Transaction Logs sitting on the same Disks/spindles). Which isn't
>>> > Best
>>> > Practice or Recommended by anyone I know (With all due respect).
>>> >
>>> > You may also consider running Microsoft Best Practice Analyser to see
>>> > if
>>> > it
>>> > gives you any recommendations. Perhaps you have write cache enabled
>>> > on
>>> > the
>>> > drives?
>>> >
>>> > I would also recommend you run performance monitor to see what the
>>> > bottleneck is before you spend money.
>>> >
>>> > Here is a link you may want to check out for Troubleshooting
>>> > Performance
>>> > with Exchange.
>>> >
>>> > http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/e2k3perf.mspx
>>> >
>>> > --
>>> > Jonathan
>>> > No Warrenties Implied, Did you do a FULL backup today??????
>>> >
>>> >
>>> >
>>> >
>>> > "Asher_N" wrote:
>>> >
>>> >> Another one of my pet peeves. Similar to the old, 'the arrow will
>>> >> never
>>> >> reach the target' problem. While your number are sound, any decent
>>> >> RAID
>>> >> controller has on-board cache. writes are cached and delayed until
>>> >> read
>>> >> are satisfied.
>>> >>
>>> >> So while the theorical problem points to fsaster writes with mirror,
>>> >> if
>>> >> you use the same drives and same RAID controler to do R1 or R5, you
>>> >> should get similar results. Also, 150 users is not much. I run 100
>>> >> users
>>> >> on a single 5 drive partitioned as 2 logical drives (OS, Stores and
>>> >> logs). The performance is more than adequate.
>>> >>
>>> >> "John Fullbright" wrote in
>>> >> news:#eQkfp6FGHA.3936@TK2MSFTNGP12.phx.gbl:
>>> >>
>>> >> > I wouldn't do that, "three disks in RAID 5 for the Database .STM
>>> >> > and
>>> >> > .EDB Files it would be 7 Disks.", that is.
>>> >> >
>>> >> > Let's assume a 3:1 read/write ratio for tha deatabases. We'll also
>>> >> > assume 10K RPM SCSI disks which are capable of ~85 20ms IOPS.
>>> >> >
>>> >> > Where P is the performance of a single spindle and N' is the number
>>> >> > of
>>> >> > data spindles in the set:
>>> >> >
>>> >> > RAID 5 Read performance = P*N' = 85*2 = 170
>>> >> > RAID 5 Write performance =P*N'/4 = 85*2/4 = 42.5
>>> >> >
>>> >> > And we apply the read/write raitio which is required when read and
>>> >> > write performance are asymmetrical. With a 3:1 ratio we are 75%
>>> >> > read
>>> >> > and 25% write:
>>> >> >
>>> >> > Performance = 170*.75 + 42.5*.25 =127.5+10.625 =138.125 IOPS
>>> >> >
>>> >> > If you were to buy bigger drives and go with just a mirror:
>>> >> >
>>> >> > RAID 1 Read performance = P*N = 170
>>> >> > RAID 1 Write performance = P*N/2 = 85
>>> >> >
>>> >> > Again we apply the read/write ratio:
>>> >> >
>>> >> > 170*.75 + 85*.25 = 127.5 + 21.3125 = 148.8125 IOPS
>>> >> >
>>> >> > My recommendation - Three mirrors for a total of 6 spindles. This
>>> >> > is
>>> >> > about 8% better overall performance on the database array with 1
>>> >> > fewer
>>> >> > spindles. The write performance is a 100% improvement.
>>> >> >
>>> >> > Do the math.
>>> >> >
>>> >> > John
>>> >> >
>>> >> >
>>> >> >
>>> >> >
>>> >> >
>>> >> > "Jonathan Norris" wrote
>>> >> > in
>>> >> > message news:D662866D-5331-4A4A-8B48-7EF998ABE340@microsoft.com...
>>> >> >> Actually if i were to use two Disks for the OS (Raid 1) and two
>>> >> >> Disks
>>> >> >> for Transaction Logs (Raid 1) and three disks in RAID 5 for the
>>> >> >> Database .STM and
>>> >> >> .EDB Files it would be 7 Disks.
>>> >> >>
>>> >> >> Also keep in mind by having everything on the same spindles you
>>> >> >> will
>>> >> >> have an
>>> >> >> I/O bottleneck.
>>> >> >>
>>> >> >> OS/Page File is Read and Write intensive
>>> >> >> Transaction Logs are write intensive during normal operation and
>>> >> >> during recovery its read intensive.
>>> >> >> DB File .stm and .edb are both Read and Write Intensive.
>>> >> >>
>>> >> >>
>>> >> >> --
>>> >> >> Jonathan
>>> >> >> No Warrenties Implied, Did you do a FULL backup today??????
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >> "Ryan Burrus" wrote:
>>> >> >>
>>> >> >>> I am trying to find some information as to the recommended
>>> >> >>> hardware
>>> >> >>> configuration for Exchange Server 2003. We only have about 150
>>> >> >>> users and all
>>> >> >>> access it using Outlook 2003. I have read a few things from
>>> >> >>> Microsoft that
>>> >> >>> says the Exchange data files and the page file should be on
>>> >> >>> separate
>>> >> >>> drives.
>>> >> >>> Also that the Exchange data files and the Windows system files
>>> >> >>> should be on
>>> >> >>> separate drives. Also that the Exchange transaction logs should
>>> >> >>> be
>>> >> >>> on a separate drive than the Exchange data files. This would
>>> >> >>> equate
>>> >> >>> to 4 drives
>>> >> >>> which doesn't sound right to me. Does anyone know what the
>>> >> >>> recommended scenario would be for an organization our size? Any
>>> >> >>> advice would be much appreciated. Thanks!!
>>> >> >
>>> >> >
>>> >> >
>>> >>
>>> >>
>>>
>>>
>>>
>
>
date: Thu, 19 Jan 2006 14:54:27 -0500
author: John Fullbright
Re: Hardware Recommendations
Ah. That'd be it. I think it was a mistake on my part to use a different
caching/controller thought. Must be using the wrong devices :)
Al
"John Fullbright" wrote in message
news:uWejVHTHGHA.1728@TK2MSFTNGP09.phx.gbl...
> Al Mulnick. Wow, another blast from the past. Still hangn' with Jim at
> Bubba's Bait & Tackle?
>
> Anyway:
>
> P = Performance of a single spindle in IOPS
> N = Number of spindles in the RAID set.
> N' = Number of data spindles in the RAID set (N - parity spindles)
>
> RAID 5 performance calculations:
>
> Write performance = P*N'/4
> Read performance = P*N'
>
> (depending on the caching architecture implemented by the controller and
> the amount of write cache, in some cases write performance can be
> expressed P*N'/3)
>
> RAID 1/10/0+1 performance caluclations:
>
> Write performance = P*N/2
> Read performance = P*N
>
> RAID 4/RAID DP (as implemented by Network Appliance) performance
> calculations
>
> Write performance = P*N'
> Read performance = P*N'
>
> For a 3 drive RAID 5 set, N' would be 2. The set size would be three and
> the other spindle is parity.
>
> "Al Mulnick" wrote in message
> news:uqljpFKHGHA.1628@TK2MSFTNGP12.phx.gbl...
>> "Interesting enough they don't come out and say RAID-1 is suitable for
>> the
>> DBs in any part of the article.
>>
>> Perhaps MS should write another article for the SMB segment. Since Raid
>> 0+1
>> isn't a cost effective solution for 150 users its open for debate.
>>
>> I am not disagreeing with you that RAID 1 provides more performance than
>> RAID 5."
>>
>> Jonathan, I see that you came to the conclusion that RAID 1 was pretty
>> good in this case. Since you're on the learning track, can I suggest
>> that you rethink the RAID concepts as you've come to know them? Also,
>> rethink the way that controllers deal with RAID. Cache is fine but keep
>> in mind that Exchange is cache-unfriendly.
>>
>> Also, when you read those documents that folks write, keep in mind they
>> write them for the general audience. Although I've personally seen some
>> of the documents for the SMB segment (and they're pretty good if I say so
>> myself), don't expect Microsoft to publish papers that say that RAID 1 is
>> recommended for Exchange. Will it work? Yep. Is it as good as RAID 5?
>> For some types of operations, yes. For others, no. If you have read
>> heavy applications, you'll get better performance from a RAID 5 than you
>> will from RAID 1. Write performance would be less. Both would have the
>> same type of fault tolerance. I throw this out there hopefully to be
>> helpful. Besides, you wouldn't have to count on tools from Rangers that
>> way ;)
>>
>> Johnathan (?), I'm surprised to see in your calculation a number of 2 for
>> the number of spindles in the data set. "RAID 5 Read performance = P*N'
>> = 85*2 = 170" Can you expand on that? I see two, but it's a raid5 set
>> meaning you have three possible targets. Are you saying that only two
>> are valid for both read and writes? Traditionally, I would have expected
>> 3 for the write and 2 for the read to account for the parity penalty.
>> Curious. I realize you're the guy to know, but I'm interested in knowing
>> why that number is used vs. another.
>>
>> Cheers,
>> Al
>>
>>
>> "Jonathan Norris" wrote in
>> message news:8EA67CFF-43E2-41FB-B0C1-76BF87504AD0@microsoft.com...
>>>I got this from
>>> http://www.microsoft.com/technet/prodtechnol/exchange/guides/StoragePerformance/1c471676-2312-4ffe-adf2-15a9cfd529c4.mspx
>>>
>>> The RAID solution you use should be based on the cost and performance
>>> trade-offs that are appropriate for your company. As a result, in many
>>> cases,
>>> more than one type of RAID solution may be recommended for a particular
>>> data
>>> storage requirement. General recommendations are as follows:
>>>
>>> . Use Raid-1+0 for the transaction log file volumes, database volumes,
>>> and
>>> SMTP queues
>>>
>>> . Use Raid-1 for the transaction log file volumes, SMTP queues, and MTA
>>> queues.
>>>
>>>
>>> . In general, Raid-5 does not provide the best trade-off between
>>> reliability/availability and performance.
>>>
>>> . Raid-0 is never recommended.
>>>
>>> The only concern I have with running RAID 1 is if you do loose a disk
>>> you
>>> will have to down the server to rebuild it unless the RAID controller
>>> supports hot swap, then again if you loose the disk in the Transaction
>>> Logs
>>> or OS your just as dead in the water.
>>>
>>> Interesting enough they don't come out and say RAID-1 is suitable for
>>> the
>>> DBs in any part of the article.
>>>
>>> Perhaps MS should write another article for the SMB segment. Since Raid
>>> 0+1
>>> isn't a cost effective solution for 150 users its open for debate.
>>>
>>> I am not disagreeing with you that RAID 1 provides more performance than
>>> RAID 5.
>>>
>>> However running the numbers... RAID 5 will support the number of users
>>> and
>>> the IOPS required to support this many users. But RAID 1 would use less
>>> disks and provide more performance.
>>>
>>> So your solution would be better in this case. I spend so little time
>>> in
>>> the SMB segment that I never implement this scenerio. I stand
>>> corrected:)
>>>
>>> Here is another article for the person who posted the question,
>>>
>>> http://www.microsoft.com/technet/prodtechnol/exchange/guides/E2k3HighAvGuide/c1804c14-c099-4207-b6b9-de5bda972b76.mspx
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> --
>>> Jonathan
>>> No Warrenties Implied, Did you do a FULL backup today??????
>>>
>>>
>>>
>>>
>>> "John Fullbright" wrote:
>>>
>>>> I don't understand what you're talking about when you say " but I
>>>> would
>>>> expect to be given my pink slip if I recommended it as a customer
>>>> solution
>>>> for a Database due to
>>>> the lack of redundancy and high risk we would assume."
>>>>
>>>> RAID 1 is 100% redundant, it's mirroring..
>>>>
>>>> The Microsoft Whitepaper "Disk Subsystem Performance Analysis for
>>>> Windows"
>>>> gives a set of criteria where it is appropriate to use RAID 5.
>>>> http://download.microsoft.com/download/e/b/a/eba1050f-a31d-436b-9281-92cdfeae4b45/subsys_perf.doc#_Toc72126973
>>>> Quoting from the paper:
>>>>
>>>> "RAID-5 might be the right choice for a write-heavy workload if the
>>>> following is true:
>>>>
>>>> The workload consists of large requests in comparison to the
>>>> array's
>>>> stripe unit size (that is, they constitute several complete stripes of
>>>> data
>>>> on average), or the controller has a write-back cache, allowing writes
>>>> to be
>>>> delayed and coalesced into full stripes, and
>>>>
>>>> The workload is largely sequential (in terms of LBNs), and
>>>>
>>>> The array controller can detect and optimize for full stripe
>>>> writes, and
>>>>
>>>> The number of spindles is reasonable, as described in "Rules
>>>> of
>>>> Thumb" later in this paper, and
>>>>
>>>> Cost is a serious issue"
>>>>
>>>>
>>>> Exchange databases are a random workload that uses a 4K (small
>>>> requests) IO
>>>> size. A three drive RAID 5 set uses more spindes at higher cost and
>>>> provides lower performance than a mirror. I don't see where any of the
>>>> criteria are met for a RAID 5 recommendation.
>>>>
>>>> The same paper discusses the characteristics of RAID 1/0+1/10:
>>>>
>>>> " Performance:
>>>> For write requests, both parts of a mirror must be updated,
>>>> reducing
>>>> the availa |
