Treadstone Seventy One: Optimizing Storage for Exchange Server 2003

Disk subsystem bottlenecks cause more performance problems than server-side CPU or RAM deficiencies, and a poorly designed disk subsystem can leave your organization vulnerable to hardware malfunctions. Specifically, your disk subsystem is performing poorly if it is experiencing:

Average read and write latencies over 20 ms.
Latency spikes over 50 ms that last for more than a few seconds.

Diskpart:

Español:
Esta tool de MS nos permite alinear las pista de los discos a las piastas de los sectores organizando de forma mas eficiente cada pista del disco y la infoamcion en el.
Windows 2000/2003 lee blques de disco en 4kb.

English:
the default starting sector for disks that have more than 63 sectors per track is the 64th sector. Because Windows will read blocks of 4 KB (8 sectors), one out of every eight blocks of data written to your disk will span two disk tracks (assuming 64 sectors per track). DiskPar can increase disk performance as much as 20 percent.

Ejemplo (EG):

1 sector---------512bt

windows lee o guarda (I/O) bloques de 4kb (8 sectores)(512*8=4KB)

|--|--|--|--|--|--|--|--| 1er bloke de 4k

|--|--|--|--|--|--|--|--| 2do bloke de 4k

|--|--|--|--|--|--|--|--| 3er bloke de 4k

|--|--|--|--|--|--|--|--|

|--|--|--|--|--|--|--|--|

|--|--|--|--|--|--|--|--|

|--|--|--|--|--|--|--|--|

|--|--|--|--|--|--|--|00| 8° bloke de 4k (00 es un sector no disponible usado por el MBR)

Total = 1 pista de 64sectores (8 sectotes (4kb) cada 8 blokes = 64 sectores)

Con disk part ustilizamos este sector optimizando en un 20% el disco.

Antes de diskpart (before diskpart):

Despues de diskpart (after diskpart):

Formatear el disco en blokes de 4kb.

Mas info:
http://www.msexchange.org/tutorials/disk-geometry.html

For example, if you have a computer running Exchange 2003 that contains one storage group with five databases, you should configure the following separate, physical RAID arrays:

C:\ - System volume, operating system, Exchange system files - RAID-1 (direct-attached storage, not SAN)
D:\ - Page file - RAID-1 (direct-attached storage, not SAN)
E:\ - SMTP and MTA queues - RAID-1+0 (SAN)
F:\ - Log files from storage group 1 - RAID-1 (SAN)
G:\ - Databases from storage group 1 - RAID-1+0 (SAN)

http://technet.microsoft.com/en-us/library/bb125079%28EXCHG.65%29.aspx

Como ver la configuracion actual de un disco:

fsutil fsinfo ntfsinfo :

Diferencias con exchange 2007:

Exch07 al ser una arquitectura x64, windows administra en una forma muy distinta la utilizacion de memoria y el use de disco, formatear los discos de base en blokes de 64k es un factor de mejora en la performance en este caso.

http://technet.microsoft.com/en-us/library/bb124518%28EXCHG.80%29.aspx

More Information:

Suppress Tuning and Use Thread Count : Automatic tuning works effectively only when you choose to host individual database and log paths on distinct disk drives (separate LUN). Jetstress may frequently fail the automatic tuning if you decide to host many storage groups on the same disk drive. We recommend that you suppress tuning and use a static number of threads for frequent automatic tuning failures.

Suppress Tuning and Use Thread Count - Jetstress may frequently fail the automatic tuning if you decide to host many storage groups on the same disk drive, so now there’s an option to suppress tuning

Configuring Jetstress Relative to Performance of a Specific Disk Subsystem

It is important to appropriately configure Jetstress in relation to the performance of the specific disk subsystem. This table provides guidelines for the number of threads for the log disk writes per second and the database disk transfers per second.

These threads are applied on a per storage group basis. The more storage groups you have, the more I/O you will have with a constant thread count.

Threads	Log Disk Writes/sec	DB Disk Transfers/sec
1	33	107
2	63	195
3	105	308
4	150	400
5	190	535
10	390	1050
20	500	1900

At a certain point, adding more threads does not necessarily increase disk throughput. It may only increase the I/O latency. We recommend that you start with fewer threads and increase them over different tests as the disk subsystem shows it can handle the load. You can determine how well the disk subsystem responds by reviewing the test results.

Analyzing Results

After the test is completed, the performance data is analyzed and reported in a summary report. Results will be saved to Performance_(DateTime).html file. All the performance counters collected will be gathered in a counter log file named Performance_(DateTime).blg that you can use for some more advanced analysis.

Consider the following guidelines when examining the data collected.

Table 1: Guidelines for examining Jetstress 2007 analysis reports
Performance Counter Instance	Guidelines for Performance Test	Guidelines for Stress Test
Database Avg. Disk sec/Read	The average value should be less than 20 ms (.20) and the maximum values should be less than 50 ms.	The maximum value should be less than 100 ms.
Log Avg. Disk sec/Write	Log disk writes are sequential, so average write latencies should be less than 10 ms, with a maximum of no more than 50 ms.	The maximum value should be no more than 100 ms.
%Processor Time	Average should be less than 80% and the maximum should be less than 90%.	Same as for Performance test.
Available Mbytes (32-bit Windows Server)	Minimum should be no less than 50 MB.	Same as for Performance test.
Free System Page Table Entries (32-bit Windows Server)	Minimum should be no less than 5000.	Same as for Performance test.
Transition Pages Repurposed/sec (Windows Server 2003)	Average should be less than 100.	Same as for Performance test.
Pages/sec (Windows 2000 Server)	Average should be less than 100.	Same as for Performance test.
Pool Nonpaged Bytes (32-bit Windows Server)	Maximum should be less than 75 MB.	Same as for Performance test.
Pool Pages Byes (32-bit Windows Server)	Maximum should be less than 180 MB.	Same as for Performance test.
Database Page Fault Stalls/sec.	Maximum should be less than 1.0.	Same as for Performance test.