|
|
|
date: 30 May 2006 11:45:02 -0700,
group: microsoft.public.exchange.clustering
back
Re: Cluster Sizing and Configuration
You haven't said anything about the load your users will generate, however
with 2451 MAPI users and 350 Blackberry users, your storage is undersized.
In terms of performance, RAID 10 is going to give you the higher numbers.
Consider:
For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
ratio is 2:1 or there about.
Let's assume 1.0 IOPS per MAPI user
For blackberry, experience shows you can add .5 IOPS per blackberry user.
Blackberry opens an addition uncached session per user. You can observe
this behavior with the exmon utility.
Our IOPS requirement is
2626 IOPS for database
263 IOPS for Logs
500 IOPS for SMTP
15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
RAID 10 read performance = P*N
RAID 10 write performance = P*N/2
WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
physical seperation of the log and database spindles, they need to be in
seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
spindles for everything else.
read = 130*10 =1300 IOPS
write = 130*10/2 = 650 IOPS
Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
You're about 2000 IOPS short. The users will definitely notice that. Even
if I used an unrealisticly rosey number like .5 IOPS per user, you're still
a thousand IOPS short. This does not include any overhead for backend
processes like backups or replication. Your actual situation is worse. You
neead to add a couple more shelves of disk.
"David B" wrote in message
news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
>I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
> an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
> below. I need to make sure that this hardware will handle my needs:
>
> Current Hardware (Stand Alone Servers)
> Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
> (RAID5)
>
> New Hardware (Active/Passive Cluster)
> Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
> (2xRAID1) internal
> One EMC Clarion CX700 DAE with a 15x146GB 15K drives
>
> Current MAPI Users 2451
> Current BlackBerry's 350
>
> 2000 users have 35MB mailbox limits
> 450 have no limits but stay under 500MB
>
> First Server will be a Stand alone Exchange 2003 swing server for SRS
> and ADC.
>
> Question:
> 1. Given this information can a two PROC server handle all of these
> users or should I upgrade to Quad PROCs now? Remember Active/Passive
> is our goal.
> 2. How many Storage Groups and Databases are recommended given 15
> disks and two hour recovery window?
> 3. Really stupid question. When running perfmon on the old system and
> I capture the % Processor Time do I use the "Average" number and if so
> do I convert it to decimal format before using it in the Microsoft
> Formulas or leave it alone.
>
> Much thanks in advance
>
date: Tue, 30 May 2006 15:19:44 -0700
author: John Fullbright [MVP] fjohn@donotspamnetappdotcom
Re: Cluster Sizing and Configuration
David, you can aditional test your Hardware using Jetstress and ESP, after
the use of Diskpar.
You can check these articles also:
CPU and memory scalability for Exchange Server 2003 and for Exchange 2000
Server - http://support.microsoft.com/default.aspx?scid=827281
Performance Benchmarks for Computers Running Exchange Server 2003 -
http://www.microsoft.com/exchange/evaluation/performance/default.mspx
How to choose server hardware for Exchange 2003 that can be effectively
re-utilized for Exchange 12? -
http://blogs.technet.com/exchange/archive/2006/03/13/421914.aspx
“Optimizing Storage For Exchange Server 2003”
http://www.microsoft.com/technet/prodtechnol/exchange/2003/library/optimizestorage.mspx
"John Fullbright [MVP]" wrote:
> You haven't said anything about the load your users will generate, however
> with 2451 MAPI users and 350 Blackberry users, your storage is undersized.
> In terms of performance, RAID 10 is going to give you the higher numbers.
> Consider:
>
> For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
> ratio is 2:1 or there about.
> Let's assume 1.0 IOPS per MAPI user
> For blackberry, experience shows you can add .5 IOPS per blackberry user.
> Blackberry opens an addition uncached session per user. You can observe
> this behavior with the exmon utility.
>
> Our IOPS requirement is
>
> 2626 IOPS for database
> 263 IOPS for Logs
> 500 IOPS for SMTP
>
>
> 15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
>
> RAID 10 read performance = P*N
> RAID 10 write performance = P*N/2
>
> WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
> physical seperation of the log and database spindles, they need to be in
> seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
> spindles for everything else.
>
> read = 130*10 =1300 IOPS
> write = 130*10/2 = 650 IOPS
>
> Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
> You're about 2000 IOPS short. The users will definitely notice that. Even
> if I used an unrealisticly rosey number like .5 IOPS per user, you're still
> a thousand IOPS short. This does not include any overhead for backend
> processes like backups or replication. Your actual situation is worse. You
> neead to add a couple more shelves of disk.
>
>
>
>
>
>
> "David B" wrote in message
> news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
> >I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
> > an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
> > below. I need to make sure that this hardware will handle my needs:
> >
> > Current Hardware (Stand Alone Servers)
> > Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
> > (RAID5)
> >
> > New Hardware (Active/Passive Cluster)
> > Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
> > (2xRAID1) internal
> > One EMC Clarion CX700 DAE with a 15x146GB 15K drives
> >
> > Current MAPI Users 2451
> > Current BlackBerry's 350
> >
> > 2000 users have 35MB mailbox limits
> > 450 have no limits but stay under 500MB
> >
> > First Server will be a Stand alone Exchange 2003 swing server for SRS
> > and ADC.
> >
> > Question:
> > 1. Given this information can a two PROC server handle all of these
> > users or should I upgrade to Quad PROCs now? Remember Active/Passive
> > is our goal.
> > 2. How many Storage Groups and Databases are recommended given 15
> > disks and two hour recovery window?
> > 3. Really stupid question. When running perfmon on the old system and
> > I capture the % Processor Time do I use the "Average" number and if so
> > do I convert it to decimal format before using it in the Microsoft
> > Formulas or leave it alone.
> >
> > Much thanks in advance
> >
>
>
>
date: Wed, 31 May 2006 11:53:02 -0700
author: David Figueiredo
Re: Cluster Sizing and Configuration
This article from HP provides the sizing for 3000 users via MAPI on Exchange
2003.
http://h71019.www7.hp.com/ActiveAnswers/downloads/5982-8353EN_Exchange2000to2003Migration.pdf
"John Fullbright [MVP]" <fjohn@donotspamnetappdotcom> wrote in message
news:uKd9ncDhGHA.4388@TK2MSFTNGP05.phx.gbl...
> You haven't said anything about the load your users will generate, however
> with 2451 MAPI users and 350 Blackberry users, your storage is undersized.
> In terms of performance, RAID 10 is going to give you the higher numbers.
> Consider:
>
> For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
> ratio is 2:1 or there about.
> Let's assume 1.0 IOPS per MAPI user
> For blackberry, experience shows you can add .5 IOPS per blackberry user.
> Blackberry opens an addition uncached session per user. You can observe
> this behavior with the exmon utility.
>
> Our IOPS requirement is
>
> 2626 IOPS for database
> 263 IOPS for Logs
> 500 IOPS for SMTP
>
>
> 15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
>
> RAID 10 read performance = P*N
> RAID 10 write performance = P*N/2
>
> WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
> physical seperation of the log and database spindles, they need to be in
> seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
> spindles for everything else.
>
> read = 130*10 =1300 IOPS
> write = 130*10/2 = 650 IOPS
>
> Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
> You're about 2000 IOPS short. The users will definitely notice that.
> Even if I used an unrealisticly rosey number like .5 IOPS per user, you're
> still a thousand IOPS short. This does not include any overhead for
> backend processes like backups or replication. Your actual situation is
> worse. You neead to add a couple more shelves of disk.
>
>
>
>
>
>
> "David B" wrote in message
> news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
>>I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
>> an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
>> below. I need to make sure that this hardware will handle my needs:
>>
>> Current Hardware (Stand Alone Servers)
>> Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
>> (RAID5)
>>
>> New Hardware (Active/Passive Cluster)
>> Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
>> (2xRAID1) internal
>> One EMC Clarion CX700 DAE with a 15x146GB 15K drives
>>
>> Current MAPI Users 2451
>> Current BlackBerry's 350
>>
>> 2000 users have 35MB mailbox limits
>> 450 have no limits but stay under 500MB
>>
>> First Server will be a Stand alone Exchange 2003 swing server for SRS
>> and ADC.
>>
>> Question:
>> 1. Given this information can a two PROC server handle all of these
>> users or should I upgrade to Quad PROCs now? Remember Active/Passive
>> is our goal.
>> 2. How many Storage Groups and Databases are recommended given 15
>> disks and two hour recovery window?
>> 3. Really stupid question. When running perfmon on the old system and
>> I capture the % Processor Time do I use the "Average" number and if so
>> do I convert it to decimal format before using it in the Microsoft
>> Formulas or leave it alone.
>>
>> Much thanks in advance
>>
>
>
date: Thu, 1 Jun 2006 20:50:02 +1000
author: Andrew Sword [MVP]
Re: Cluster Sizing and Configuration
With puny (by today's standards) 20mb mailboxes and presumably an IOPS/user
figure in the .25 - .5 range, 28 disks.
"Tables 3 and 4 show the total time it took to migrate mailboxes during
testing. Note the size of data being moved is approximately 20 MB per user
mailbox, as is initialized by LoadSim when using the MMB2 profile."
"The EVA3000 (2C2D) configuration included:
1.. . Twenty-eight 36-GB, 15,000-rpm disks
2.. . Two disk groups
1.. - Database disk group, 14 disks
2.. - Transaction log disk group, 14 disks "
"Andrew Sword [MVP]" wrote in message
news:uF25ikWhGHA.1856@TK2MSFTNGP03.phx.gbl...
> This article from HP provides the sizing for 3000 users via MAPI on
> Exchange 2003.
>
> http://h71019.www7.hp.com/ActiveAnswers/downloads/5982-8353EN_Exchange2000to2003Migration.pdf
>
>
>
>
> "John Fullbright [MVP]" <fjohn@donotspamnetappdotcom> wrote in message
> news:uKd9ncDhGHA.4388@TK2MSFTNGP05.phx.gbl...
>> You haven't said anything about the load your users will generate,
>> however with 2451 MAPI users and 350 Blackberry users, your storage is
>> undersized. In terms of performance, RAID 10 is going to give you the
>> higher numbers. Consider:
>>
>> For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
>> ratio is 2:1 or there about.
>> Let's assume 1.0 IOPS per MAPI user
>> For blackberry, experience shows you can add .5 IOPS per blackberry user.
>> Blackberry opens an addition uncached session per user. You can observe
>> this behavior with the exmon utility.
>>
>> Our IOPS requirement is
>>
>> 2626 IOPS for database
>> 263 IOPS for Logs
>> 500 IOPS for SMTP
>>
>>
>> 15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
>>
>> RAID 10 read performance = P*N
>> RAID 10 write performance = P*N/2
>>
>> WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
>> physical seperation of the log and database spindles, they need to be in
>> seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
>> spindles for everything else.
>>
>> read = 130*10 =1300 IOPS
>> write = 130*10/2 = 650 IOPS
>>
>> Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
>> You're about 2000 IOPS short. The users will definitely notice that.
>> Even if I used an unrealisticly rosey number like .5 IOPS per user,
>> you're still a thousand IOPS short. This does not include any overhead
>> for backend processes like backups or replication. Your actual situation
>> is worse. You neead to add a couple more shelves of disk.
>>
>>
>>
>>
>>
>>
>> "David B" wrote in message
>> news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
>>>I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
>>> an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
>>> below. I need to make sure that this hardware will handle my needs:
>>>
>>> Current Hardware (Stand Alone Servers)
>>> Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
>>> (RAID5)
>>>
>>> New Hardware (Active/Passive Cluster)
>>> Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
>>> (2xRAID1) internal
>>> One EMC Clarion CX700 DAE with a 15x146GB 15K drives
>>>
>>> Current MAPI Users 2451
>>> Current BlackBerry's 350
>>>
>>> 2000 users have 35MB mailbox limits
>>> 450 have no limits but stay under 500MB
>>>
>>> First Server will be a Stand alone Exchange 2003 swing server for SRS
>>> and ADC.
>>>
>>> Question:
>>> 1. Given this information can a two PROC server handle all of these
>>> users or should I upgrade to Quad PROCs now? Remember Active/Passive
>>> is our goal.
>>> 2. How many Storage Groups and Databases are recommended given 15
>>> disks and two hour recovery window?
>>> 3. Really stupid question. When running perfmon on the old system and
>>> I capture the % Processor Time do I use the "Average" number and if so
>>> do I convert it to decimal format before using it in the Microsoft
>>> Formulas or leave it alone.
>>>
>>> Much thanks in advance
>>>
>>
>>
>
>
date: Thu, 1 Jun 2006 08:32:23 -0700
author: John Fullbright [MVP] fjohn@donotspamnetappdotcom
Re: Cluster Sizing and Configuration
Can you show how you arrived at the IOPS requirements?
"John Fullbright [MVP]" wrote:
> You haven't said anything about the load your users will generate, however
> with 2451 MAPI users and 350 Blackberry users, your storage is undersized.
> In terms of performance, RAID 10 is going to give you the higher numbers.
> Consider:
>
> For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
> ratio is 2:1 or there about.
> Let's assume 1.0 IOPS per MAPI user
> For blackberry, experience shows you can add .5 IOPS per blackberry user.
> Blackberry opens an addition uncached session per user. You can observe
> this behavior with the exmon utility.
>
> Our IOPS requirement is
>
> 2626 IOPS for database
> 263 IOPS for Logs
> 500 IOPS for SMTP
>
>
> 15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
>
> RAID 10 read performance = P*N
> RAID 10 write performance = P*N/2
>
> WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
> physical seperation of the log and database spindles, they need to be in
> seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
> spindles for everything else.
>
> read = 130*10 =1300 IOPS
> write = 130*10/2 = 650 IOPS
>
> Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
> You're about 2000 IOPS short. The users will definitely notice that. Even
> if I used an unrealisticly rosey number like .5 IOPS per user, you're still
> a thousand IOPS short. This does not include any overhead for backend
> processes like backups or replication. Your actual situation is worse. You
> neead to add a couple more shelves of disk.
>
>
>
>
>
>
> "David B" wrote in message
> news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
> >I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
> > an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
> > below. I need to make sure that this hardware will handle my needs:
> >
> > Current Hardware (Stand Alone Servers)
> > Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
> > (RAID5)
> >
> > New Hardware (Active/Passive Cluster)
> > Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
> > (2xRAID1) internal
> > One EMC Clarion CX700 DAE with a 15x146GB 15K drives
> >
> > Current MAPI Users 2451
> > Current BlackBerry's 350
> >
> > 2000 users have 35MB mailbox limits
> > 450 have no limits but stay under 500MB
> >
> > First Server will be a Stand alone Exchange 2003 swing server for SRS
> > and ADC.
> >
> > Question:
> > 1. Given this information can a two PROC server handle all of these
> > users or should I upgrade to Quad PROCs now? Remember Active/Passive
> > is our goal.
> > 2. How many Storage Groups and Databases are recommended given 15
> > disks and two hour recovery window?
> > 3. Really stupid question. When running perfmon on the old system and
> > I capture the % Processor Time do I use the "Average" number and if so
> > do I convert it to decimal format before using it in the Microsoft
> > Formulas or leave it alone.
> >
> > Much thanks in advance
> >
>
>
>
date: Mon, 5 Jun 2006 09:53:01 -0700
author: SilverICE
Re: Cluster Sizing and Configuration
"Exchange 2003 with Outlook 2003 cached mode clients, the read/write
>> ratio is 2:1 or there about.
>> Let's assume 1.0 IOPS per MAPI user
>> For blackberry, experience shows you can add .5 IOPS per blackberry user.
>> Blackberry opens an addition uncached session per user. You can observe
>> this behavior with the exmon utility."
Since there is no measurement, these were the assumptions stated.
Even at the 20mb mailbox size referenced in the HP document,
http://h71019.www7.hp.com/ActiveAnswers/downloads/5982-8353EN_Exchange2000to2003Migration.pdf ,
a .5 IOPS assumtion was used (28 15K spindles). In this enviroment, David
B. states that 2000 users have a 35MB mailbox limit and 450 have no limit
but stay under 500mb. This is why I used 1 IOP/user as the assumption. In
measuring you'll find that those "unlimited" users will substantially
increase the IOPS/user ratio, so I recommend measurement. The point of the
assumptions was to demonstrate how woefully undersized the storage was.
"SilverICE" wrote in message
news:2A1E7F1D-F770-4442-83C7-D7C5CE696DF5@microsoft.com...
> Can you show how you arrived at the IOPS requirements?
>
>
> "John Fullbright [MVP]" wrote:
>
>> You haven't said anything about the load your users will generate,
>> however
>> with 2451 MAPI users and 350 Blackberry users, your storage is
>> undersized.
>> In terms of performance, RAID 10 is going to give you the higher numbers.
>> Consider:
>>
>> For Exchange 2003 with Outlook 2003 cached mode clients, the read/write
>> ratio is 2:1 or there about.
>> Let's assume 1.0 IOPS per MAPI user
>> For blackberry, experience shows you can add .5 IOPS per blackberry user.
>> Blackberry opens an addition uncached session per user. You can observe
>> this behavior with the exmon utility.
>>
>> Our IOPS requirement is
>>
>> 2626 IOPS for database
>> 263 IOPS for Logs
>> 500 IOPS for SMTP
>>
>>
>> 15K FC drives @ 130 IOPS/spindle at a target respone time of 20ms.
>>
>> RAID 10 read performance = P*N
>> RAID 10 write performance = P*N/2
>>
>> WIth 15 drives, lets assume 1 will be a hot spare. In order to maintain
>> physical seperation of the log and database spindles, they need to be in
>> seperate RAID groups. For the logs, we need 4 spindles: This leaves 10
>> spindles for everything else.
>>
>> read = 130*10 =1300 IOPS
>> write = 130*10/2 = 650 IOPS
>>
>> Apply the read/write ratio 1300*.66 + 650 *.33 = 850 + 215 = 1065 IOPS.
>> You're about 2000 IOPS short. The users will definitely notice that.
>> Even
>> if I used an unrealisticly rosey number like .5 IOPS per user, you're
>> still
>> a thousand IOPS short. This does not include any overhead for backend
>> processes like backups or replication. Your actual situation is worse.
>> You
>> neead to add a couple more shelves of disk.
>>
>>
>>
>>
>>
>>
>> "David B" wrote in message
>> news:1149014702.564619.79760@i40g2000cwc.googlegroups.com...
>> >I currently have 2,451 users (Exchange1=1,325 and Exchange2=1,126) in
>> > an Exchange 5.5 Org. NT4 domain. I have new hardware and one DAE, see
>> > below. I need to make sure that this hardware will handle my needs:
>> >
>> > Current Hardware (Stand Alone Servers)
>> > Two Dell PowerEdge 6400 with 4x700Mhz, 1GB RAM, 150GB of Storage
>> > (RAID5)
>> >
>> > New Hardware (Active/Passive Cluster)
>> > Two Dell PowerEdge 6800 with 2x3.0Ghx, 4GB RAM, 300GB of Storage
>> > (2xRAID1) internal
>> > One EMC Clarion CX700 DAE with a 15x146GB 15K drives
>> >
>> > Current MAPI Users 2451
>> > Current BlackBerry's 350
>> >
>> > 2000 users have 35MB mailbox limits
>> > 450 have no limits but stay under 500MB
>> >
>> > First Server will be a Stand alone Exchange 2003 swing server for SRS
>> > and ADC.
>> >
>> > Question:
>> > 1. Given this information can a two PROC server handle all of these
>> > users or should I upgrade to Quad PROCs now? Remember Active/Passive
>> > is our goal.
>> > 2. How many Storage Groups and Databases are recommended given 15
>> > disks and two hour recovery window?
>> > 3. Really stupid question. When running perfmon on the old system and
>> > I capture the % Processor Time do I use the "Average" number and if so
>> > do I convert it to decimal format before using it in the Microsoft
>> > Formulas or leave it alone.
>> >
>> > Much thanks in advance
>> >
>>
>>
>>
date: Mon, 5 Jun 2006 10:36:25 -0700
author: John Fullbright [MVP] fjohn@donotspamnetappdotcom
|
|
