Dell PowerVault MD3000i Quick start guide

ARRAY TUNING
BEST PRACTICES
dell.com/PowerVault
POWERVAULT
MD3000 AND
MD3000i

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page2
DISCLAIMER
THIS WHITE PAPER IS FOR INFORMATIONAL PURPOSES ONLY, AND MAY
CONTAIN TYPOGRAPHICAL ERRORS AND TECHNICAL INACCURACIES.
THE CONTENT IS PROVIDED AS IS, WITHOUT EXPRESS OR IMPLIED
WARRANTIES OF ANY KIND.
For more information, contact Dell™.
Information in this document is subject to change without notice.
http://www.dell.com

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page3
Table of Contents
1AUDIENCEANDSCOPE .............................................................................................................................. 4
2PERFORMANCETUNINGOVERVIEW .......................................................................................................... 4
2.1COMPONENTSTHATINFLUENCESTORAGEPERFORMANCE ...........................................................................................4
2.2BASICAPPROACHTOPERFORMANCETUNING ............................................................................................................4
3APPLICATIONSOFTWARECONSIDERATIONS.............................................................................................. 5
4CONFIGURINGTHEMD3000/MD3000I ...................................................................................................... 6
4.1DETERMININGTHEBESTRAIDLEVEL ........................................................................................................................6
4.1.1SelectingaRAIDLevel‐HighWriteMixScenario..................................................................................8
4.1.2SelectingaRAIDLevel‐LowWriteMixScenario...................................................................................8
4.2CHOOSINGTHENUMBEROFDRIVESINADISKGROUP.................................................................................................8
4.3VIRTUALDISKLOCATIONANDCAPACITY ...................................................................................................................9
4.4VIRTUALDISKOWNERSHIP...................................................................................................................................10
4.5CALCULATINGOPTIMALSEGMENTANDSTRIPESIZE ..................................................................................................10
4.6CACHESETTINGS ................................................................................................................................................12
4.6.1SettingtheVirtualDisk‐SpecificWriteCacheandWriteCacheMirroring...........................................12
4.6.2SettingtheVirtualDisk‐SpecificReadCachePre‐fetch........................................................................12
4.6.3SettingtheStorageArrayCacheBlockSize .........................................................................................13
4.7TUNINGUSINGARRAYPERFORMANCEDATA ...........................................................................................................14
4.7.1CollectingPerformanceStatistics ........................................................................................................14
4.7.2RAIDLevel ............................................................................................................................................14
4.7.3I/ODistribution ....................................................................................................................................15
4.7.4StripeSize.............................................................................................................................................16
4.7.5WriteAlgorithmData ..........................................................................................................................19
4.8USINGTHECLIPERFORMANCEMONITOR ...............................................................................................................21
4.9OTHERARRAYCONSIDERATIONS ...........................................................................................................................22
4.9.1GlobalMediaScanRate.......................................................................................................................22
4.9.2SettingtheVirtualDisk‐SpecificMediaScan .......................................................................................22
4.10PREMIUMFEATUREPERFORMANCE ..................................................................................................................22
4.10.1GettingOptimalPerformancefromSnapshot.................................................................................22
4.10.2GettingOptimalPerformancefromVirtualDiskCopy ....................................................................22
5CONSIDERINGTHEHOSTSERVER(S)..........................................................................................................23
5.1HOSTHARDWAREPLATFORM ...............................................................................................................................23
5.1.1ConsideringtheServerHardwareArchitecture....................................................................................23
5.1.2SharingBandwidthontheDell™MD3000iwithMultipleNICs ...........................................................23
5.1.3SharingBandwidthwithMultipleSASHBAs ........................................................................................24
5.2CONSIDERINGTHESYSTEMSOFTWARE ...................................................................................................................25
5.2.1BufferingtheI/O ..................................................................................................................................25
5.2.2AligningHostI/OwithRAIDStriping....................................................................................................25
APPENDIXA:OBTAININGADDITIONALPERFORMANCETOOLS ..........................................................................27
APPENDIXB:SYSTEMTROUBLESHOOTING .......................................................................................................28
APPENDIXC:REFERENCES ................................................................................................................................29
APPENDIXD:GLOSSARYOFTERMS ...................................................................................................................30

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page4
1 Audience and Scope
This document is intended to guide MD3000 and MD3000i customers through
the advanced processes behind tuning their storage array to best suit their
individual needs. Presented within are the best practices that must be observed
in performance tuning a first generation firmware (06.XX.XX.XX) and second
generation firmware (07.XX.XX.XX) storage array. For further detail on
determining Firmware Generation of a MD3000 or MD3000i storage array,
please view the Dell™ Users Guide at http://support.dell.com/manuals.
2 Performance Tuning Overview
The challenge of storage performance tuning is to understand and control the
interacting components (listed below) while accurately measuring the application
performance. Because the performance of the storage array accounts for only a
portion of the overall application performance, tuning must be performed taking
into account the input/output (I/O) characteristics of the application and all of the
components involved in the data path, such as the SAS HBA, iSCSI initiator,
network switch and host operating system settings.
With multiple aspects to consider, the task of performance tuning to maximize
performance for even one application can seem formidable. Tuning the system to
maximize performance of multiple applications that may share a single storage
array can seem even more formidable. To reduce the complexity of tuning, Dell™
storage systems feature performance monitoring and flexible tuning controls that
can be accessed through the Modular Disk Storage Manager (MDSM).
2.1 Components That Influence Storage Performance
This white paper provides an overall approach to tuning I/O performance and
also provides specific guidelines for using the storage array tuning controls.
These recommendations begin with an overall analysis of the elements that
determine I/O performance:
•Storage array
•Application software
•Server platform (hardware, operating system, volume managers,
device drivers)
•Network (for MD3000i only)
2.2 Basic Approach to Performance Tuning
The first principles of I/O performance tuning include the following question:
What should be the performance for my system?

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page5
The answers include:
•“It depends…” There are no absolute answers. Each environment is
unique and the correct settings depend on the unique goals,
configuration, and demands for the specific environment.
•“Actual mileage may vary.” Results vary widely because conditions
vary widely.
The answers to this question suggest the following basic approach to
performance tuning:
1– Configure and test
2– Measure
3– Adjust as required
The performance monitoring features in all MD3000 and MD3000i storage
systems and the controls for tuning make them ideally suited for this iterative
process. The first step in tuning is to establish a baseline of existing
performance. When generating a performance baseline, it is ideal to use a
workload that is as similar to the intended final use of the storage solution. This
can be as simple as the real application or a SQL Replay with a system
performance monitor (perfmon or sysstat/iostat as well as the CLI and state-
capture performance monitoring) or a synthetic benchmark package that closely
imitates the expected I/O profile (Iometer, IOZone, Bonnie). By comparing the
baseline data to the estimated needs and capability of the configuration, the user
can effectively tune an MD3000 or MD3000i storage array. This white paper
provides recommendations for this important first step as well as tuning tweaks to
get the full capabilities from the MD3000 and MD3000i storage systems.
3 Application Software Considerations
In understanding the I/O characteristics of intended applications using the
storage in a manner as close to the expected run-time is required to determine
the optimal storage and system configuration, and is critical to tuning the overall
solution. This includes, but is not limited to the following:
•Number of discrete I/O sources interacting with the solution
•Randomness of data access by the primary I/O source(s)
•Mean size of typical I/O; this is usually divided into the following three
categories:
oLarge Block ( ≥256KiB ) transfer size
oMedium Block (≥32KiB and < 256KiB) transfer sizes
oSmall Block ( < 32KiB) transfer sizes
•Burstiness of I/O pattern, i.e. the mean duty-cycle of I/O to the storage
array

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page6
•Profile of mean I/O direction; this is usually the ratio of Reads to Writes
4 Configuring the MD3000/MD3000i
There are two ways to configure the MD3000 and MD3000i storage systems.
The most common and easiest method is using the MDSM. The MDSM allows
automatic configuration settings that provide reasonable configurations with little
knowledge of performance tuning required.
A manual configuration option is also available using the command line interface
(CLI), providing more flexibility but requiring more knowledge of performance
requirements.
For a link to the MDSM and CLI guides, see Appendix C: References.
4.1 Determining the Best Raid Level
The first step involved in tuning an MD3000 or MD3000i external storage array is
to determine the RAID level most appropriate for the solutions given the
application. In the proceeding document, please note that RAID 0 is excluded
from most information in this paper due to the lack of data protection. This does
not mean the use of RAID 0 is undesired, just that it should only be used for non-
critical data. In general, RAID 0 generally provides better performance than RAID
1/10, 5 or 6. Additionally, RAID 6 is not always referenced specifically; most
comments applicable to tuning RAID 5 are directly applicable to RAID 6 unless
otherwise noted. In situations where the increased fault tolerance provided by
RAID 6 are desired, please note that a performance penalty is observed when
compared directly to RAID 5 due to the additional parity calculation and the extra
physical disk required for implementation.
An important consideration when determining the appropriate RAID level is the
physical disk cost required by a RAID level. Physical disk cost, is the number of
physical drives worth of capacity that are sacrificed to provided the desired data
integrity level. The physical disk cost of each RAID level is different, and may
impact the decision for which RAID level is most appropriate for a chosen
environment. RAID 0, having no level of redundancy has a physical disk cost of
zero. RAID 1/10 has the highest disk-cost in disk groups containing more than 2
drives. Half of the physical drives in a RAID 1/10 are always consumed for the
mirror. RAID 5 has a fixed 1-physical disk per disk-group cost, i.e. with a RAID 5
set of n disks, only n-1 worth of capacity is available. Similarly, RAID 6 has a
fixed two physical disks per disk-group cost, or n-2. In RAID 5 and 6, these extra
drives account for the space required to maintain the levels of parity information
for each stripe.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page7
Physical disk cost is not the only factor that influences the decision on which
RAID level is most appropriate for a given application. The performance of a
chosen RAID level is heavily interdependent on characteristics of the I/O pattern
as transmitted to the storage array from the host(s). With I/O patterns involving
write operations, when an I/O burst exceeds 1/3 of available cache memory in
size, it should be considered a long I/O. Long writes show the performance of a
chosen RAID level better than short writes. Short write operations can be
handled entirely in cache, and the RAID level performance effect is minimized.
As long as the write-burstiness is always lower than the cache to disk offload
rate, a choice in RAID level can be a non-issue.
In general, the following outlines which RAID levels work best in specific
circumstances:
•RAID 5 and RAID 6 works best for sequential, large I/Os (>256KiB)
•RAID 5 or RAID 1/10 for small I/Os ( <32KiB )
•For I/O sizes in between, the RAID level is dictated by other application
characteristics:
oRAID 5 and RAID 1/10 have similar characteristics for most Read
environments and sequential Writes.
oRAID 5 and RAID 6 exhibit the worst performance mostly by
random writes.
oIn random I/O applications consisting of more than 10% write
operations, RAID 1/10 provides the best performance.
Table 1 provides a summary of these points for an ideal environment. An ideal
environment consists of aligned stripes or segments reads and writes, as well as
burst I/O where cache memory and RAID Controller Modules are not overly
saturated by I/O operations.
Table 1: I/O Size and Optimal Raid Level
Significantly Random Significantly Sequential
Block Size Read Write Read Write
Small
(<32 KiB ) 1/10, 5, 6 1/10 1/10, 5, 6 1/10, 5
Medium
(Between 32
and 256 KiB)
1/10, 5, 6 1/10 1/10, 5, 6 5
Large
(>256 KiB) 1/10, 5, 6 1/10 1/10, 5, 6 5

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page8
4.1.1 Selecting a RAID Level - High Write Mix Scenario
In random I/O applications with a >10% mix of write operations and a low degree
of burstiness, RAID 1/10 provides the best overall performance for redundant
disk groups.
RAID 1/10 performance can be >20% better than RAID 5 in these environments,
but has the highest disk cost; for instance, more physical disks must be
purchased. RAID 5 provides protection and minimizes disk cost for net capacity
but is heavily impacted by the write performance overhead of parity updates.
RAID 6, while providing better protection than RAID 5 with minimal disk cost, is
more heavily impacted by write performance overhead for the double parity
calculations it requires.
In sequential I/O applications with relatively small write transfer sizes, RAID level
does not make a large difference. With medium transfer sizes, RAID 1/10 can
provide an advantage over RAID 5/6, again with a higher associated disk-cost.
In very large sequential writes, RAID 5 can perform equal or better than RAID
1/10 especially when factoring in disk cost for equivalent capacity. Additionally,
better performance is always achieved when the application or OS is capable of
buffering or coalescing writes to fill a whole segment or stripe.
4.1.2 Selecting a RAID Level - Low Write Mix Scenario
In random I/O applications with a low mix (<10%) of write operations, RAID 5
offers approximately the same performance as RAID 1/10 in small transfer sizes
but at a lower disk cost. RAID 0 provides somewhat better performance than 5
or 1/10 but offers no data protection. In larger transfer size environments, RAID
1/10 performances can be slightly better than RAID 5 but has a much higher disk
cost.
4.2 Choosing the Number of Drives in a Disk Group
When optimizing performance, there are many factors to consider, such as drive
type, capacity, and the number of drives.
The following general guidelines can be used in grouping drives in a disk group:
•Separate random and sequential workloads on different disks groups –
segregating I/O traffic to minimize sharing of disk groups between virtual
disks.
•Choose faster drives – in general, a single 15K RPM drive yields
approximately 20% more performance than 10K for mixed sequential
and random operations. Please refer to the manufacturer’s data sheets
to better determine the optimally suitable drive.
•Adding more drives to a disk group while holding the stripe size fixed
can increase I/O rate for sequential I/O, up to the point of controller
saturation – more drives mean more spindles to service I/O.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page9
•To optimize for data transfer rate, multiply the number of physical data
disks by the segment size to equal the I/O size. However, there are
always exceptions. For small/medium I/Os care should be taken to avoid
splitting up the I/O in such a way that an even smaller I/O is sent to the
disk drives. Please note that data disks do not include parity or mirror
disks used in a RAID set.
For IOPS or transaction-oriented applications, the number of drives becomes
more significant because disk drive random I/O rates are relatively low. Select a
number of drives that matches the per virtual disk group I/O rate needed to
support the application. Make sure to account for the I/Os required to implement
the data protection of the selected RAID level. Make the segment size at least as
large as the typical application I/O size.
A segment size of 128K is a reasonable starting point for most applications. In
most applications, the greater the number of drives in a disk group, the better the
average performance. The drive count of an existing disk group can be increased
using the CLI.
4.3 Virtual Disk Location and Capacity
The location of virtual disks within a disk group, the number and location of virtual
disks allocated within a disk group, and the capacity of a virtual disk are
important factors to consider when optimizing an array for performance.
When using rotating storage medium, the capacity of a virtual disk and its
location within a disk-group greatly impact achieved performance. This is
primarily due to differences in angular velocity in the outside zones. The effect of
allocating the outermost edges of a rotational storage medium to increase
performance is termed short stroking of a drive. While it is beyond the scope of
this white paper to cover the technical details involved around short-stroking,
typically the outer third of a rotational medium is the fastest, while the inner most
zones are the slowest. Short stroking can easily be accomplished by creating a
disk group consisting of a single virtual disk, which is allocated less than a third
of the total capacity. The obvious downside to short-stroking a volume is the loss
of additional usable capacity. Thus, this performance gain must be weighed
directly against the capacity loss.
In addition to the performance gains of short-stroking, the effect of drive head
seek should be taken into account when carving up a disk group into virtual
disks. Virtual disks are aligned in series within a Disk Group with the first virtual
disk starting in the fastest outer regions and progressing inwards. When taking
this into account, a disk group should be designed with as few virtual disks as
possible.
Dell™ does not recommend using more than four virtual disks or repositories per
disk group for peak performance. Additionally, where performance is critical,
isolate virtual disks to separate disk groups as possible. When multiple high-

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page10
traffic virtual disks share a disk group, even with purely sequential usage models,
the disk group I/O behavior becomes increasingly random, lowering overall
performance. Additionally, when a disk group must be shared, the most heavily
trafficked virtual disk always should be located at the beginning of a disk group.
4.4 Virtual Disk Ownership
The Dell™ MDSM can be used to automatically build and view virtual disks. It
uses optimal settings to stripe the disk group. Virtual disks are assigned to
alternating RAID controllers when they are created. This default assignation
provides a simple means for load balancing the workload of the RAID controllers.
Ownership can later be modified to balance workload according to actual usage. If
virtual disk ownership is not manually balanced, it is possible for one controller to
have the majority of the work, while the other controller is idle.
Limit the number of virtual disks in a disk group. If multiple virtual disks are in a
disk group, consider the following information:
•Consider the impact each virtual disk has on other virtual disks in the
same disk group.
•Understand the patterns of usage for each virtual disk.
•Different virtual disks have higher usage at different times of day.
Figure 1: Virtual disk ownership balance.
4.5 Calculating Optimal Segment and Stripe Size
The choice of a segment size can have a major influence on performance in both
IOPS and data transfer rate.
The term segment size refers to the amount of data written to one drive in a
virtual disk group before writing data to the next drive in the virtual disk group. A
set of contiguous segments spanning across member drives creates a stripe. For
example, in a RAID 5, 4 + 1 virtual disk group with a segment size of 128KiB, the

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page11
first 128KiB of an I/O is written to the first drive, the next 128KiB to the next drive,
and so on with a total stripe size of 512KiB. For a RAID 1, 2 + 2 virtual disk
group, 128KiB would be written to each of the two drives (and same for the
mirrored drives). If the I/O size is larger than this (the number of physical disks
multiplied by a 128KiB segment), this pattern repeats until the entire I/O is
complete.
For very large I/O requests, the optimal segment size for a RAID volume group is
one that distributes a single host I/O across all data drives within a single stripe.
The formula for maximal stripe size is as follows:
LUN segment size = Maximal I/O Size ÷number of data drives
However, the LUN segment size should be rounded up to the nearest supported
power of two value.
For RAID5 and 6, the number of data drives is equal to the number of drives in
the volume group minus 1 and 2 respectively. For example:
RAID5, 4+1 with a 64KiB segment size => (5-1) x 64KiB = 256KiB stripe
size
Optimally, this RAID group is sufficient for handling I/O requests less than or
equal to 256KiB.
For RAID1, the number of data drives is equal to the number of drives divided by
2. For example:
RAID1/10, 2+2 with a 64KiB segment size => (4-2) x 64KiB = 128KiB
stripe size
It is important to remember that depending on the application’s I/O parameters,
segment and strip size will vary.
For application profiles with small I/O requests, set the segment size large
enough to minimize the number of segments (drives in the LUN) that are
accessed to satisfy the I/O request, that is, to minimize segment boundary
crossings. Unless specified otherwise by the application, starting with the default
segment size of 128KiB is recommended.
It is imperative that the stripe size be correctly selected so that the host operating
system is always making requests properly aligned with full stripes or full
segments when possible.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page12
4.6 Cache Settings
Read-ahead cache can be configured in the MDSM and through the CLI. The
MDSM has only the defaults to work with, while the CLI can fully configure the
read-ahead cache. Additionally, the global cache block size for read and write
cache can be adjusted through the CLI.
Please refer to the Dell™ PowerVault™ Modular Disk Storage Manager CLI
Guide on the Dell™ technical support site (http://support.dell.com/manuals) for a
complete list of supported commands including the following cache specific
commands.
4.6.1 Setting the Virtual Disk-Specific Write Cache and Write Cache Mirroring
Configured through the CLI – These commands are available at the virtual disk
level.
Write Cache – Disabling write cache puts the controllers into a write-through
mode, adding additional latency while data is flushed to the disk. Except for
specific read-only environments, it is recommended that this setting stay
enabled. Write cache is automatically disabled in the case of cache battery
failure or a cache battery learn cycle.
Write Cache Mirroring – Write Cache Mirroring provides an additional level of
redundancy and fault tolerance in the MD3000 and MD3000i. As a side effect, it
reduces available physical memory and intra-controller bandwidth to perform this
operation. In select, non data-critical cases, it can be beneficial to adjust this
parameter. For normal use, Dell™ always recommends Enabling Cache
Mirroring. Cache Mirroring is automatically disabled in the event of controller
failure or when write caching is disabled.
WARNING: Data loss can occur if a RAID controller module fails while write-
caching without cache mirroring is enabled on a virtual disk.
4.6.2 Setting the Virtual Disk-Specific Read Cache Pre-fetch
Configured through the CLI – This command is available at the virtual disk level.
Read Cache Pre-fetch – The Read Cache setting can be toggled on a virtual disk
level. Disabling read pre-fetch is primarily useful in primarily small transfer size
random read environments, where pre-fetch of random data would not provide
sufficient value. However, the normal observed overhead of read pre-fetch is
negligible. For most environments, Dell™ always recommends Enabling Read
Pre-fetch.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page13
4.6.3 Setting the Storage Array Cache Block Size
Configured through the CLI – This command is available at the storage array
level and effects all virtual disks and disk-groups.
Cache Block Size – Cache Block Size refers to the way cache memory is
segmented during allocation and affects all virtual disks in an array. On the
MD3000 and MD3000i, settings of 4KiB and 16KiB are available with 4KiB being
the default. A dramatic impact on performance can occur by choosing the
correct cache block size setting specific to the system’s I/O profile. If the typical
I/O size is ≥16KiB, which is typical with sequential I/O, set the storage array
Cache Block Size to 16. For smaller (≤8KiB) I/O, especially in highly random or
transactional use cases, the default 4KiB setting is preferred. As this setting
affects all virtual disks on a storage array, changing it should be done with
attention to the I/O needs of the application.
Table 2: Storage Array Standard Configuration Specifications
MDSM GUI configuration templates
Option File System Database Multimedia CLI Options
Drive Type Selectable Selectable Selectable Selectable
RAID Level Selectable Selectable Selectable 0, 1/10, 5, 6
Segment Size 1128KiB 128KiB 256KiB
8KiB, 16KiB,
32KiB, 64KiB,
128KiB, 256KiB,
512KiB
Write cache
with mirroring Fixed to on Fixed to on Fixed to on On or Off
Read-ahead
cache On Off On On or Off
Cache Block
Size 1Array Defaults to 4KiB 4KiB, 16KiB
1The MDSM CLI Guide and SMcli application may use the legacy terminology of KB to mean
Kilobytes, or 210 bytes. Throughout this whitepaper, the SI term kibibytes are used instead.
When formulating an SMcli command the KB postfix is however still required. IEEE 1541-2002
and IEC 60027-2 standards detail units of measurement prefixes for binary multiples.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page14
4.7 Tuning Using Array Performance Data
4.7.1 Collecting Performance Statistics
The stateCaptureData.txt and performanceStatistics.csv files, which are available
through the MDSM, Support tab as part of a Technical Support Bundle, provide
valuable statistical data in an easy-to-read format. The following section shows
some sample data from the stateCaptureData.txt file and suggested
configuration recommendations based on the performance considerations
outlined in the previous section.
Other useful information is available through the array profile. Open the MDSM
and select the Support tab – View Storage Array Profile.
Before collecting performance statistics, the I/O workload under test should be
first executed. This will ensure the validity of performance statistics as part of the
measure step of proper performance tuning.
Note: Figures shown below are from using the performance tool Iometer.
4.7.2 RAID Level
The stateCaptureData.txt file provides statistics in the read and writes
percentage columns to aid in selecting most appropriate RAID level. In Figure 2,
the small reads and writes I/O percentages provide information regarding the
distribution of the I/O types in the tested workload. This data is especially helpful
when utilizing Table 1 referenced on page 7 to determining the applications
current read/write mix. The RAID Level chosen can impact the I/O performance.
Generally, RAID 1/10 has the best overall performance, with the highest physical
disk cost. Use the I/O percent distribution and the average block size from the
collected data to aid in this determination. These fields can be found in the
highlighted regions of Figure 2 and Figure 3 for First and Second Generation
Firmware respectively. It should be noted that the values in these figures are in
block notation; block size for the specific virtual disk configuration is listed in the
stateCaptureData.txt file, and is almost always 512 bytes. The average received
I/O is not the I/O size the application uses but what the host sends, so while an
application may attempt to send larger I/Os, the host’s I/O stack may coalesce or
break up I/Os as appropriate. Please see the appropriate separate OS or HBA
documentation to determine these values.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page15

4.7.3 I/O Distribution
I/O can be characterized by its distribution and pattern. The two primary factors
in determining the I/O distribution of an application are the randomness of I/O
and the direction of I/O. The randomness of I/O indicates how sequential or
random the data access is, as well as the patterning of this data access. The
direction of the I/O can be simply related to the read and write percentages of
I/O, that is, the direction I/O is taking from the storage device. I/O pattern refers
Volume 0 Attributes:
Volume Type: RAIDVolume
User Label: MyRAID10_One
...
BlockSize: 512 bytes
LargeIoSize: 4096 blocks
...
Perf. Stats: Requests Blocks Avg. Blks IO Percent
Reads 67456452 5943724625 88 71.20%
Writes 27283249 1144902648 41 28.80%
Large Reads 0 0 0 0.00%
Large Writes 0 0 0 0.00%
Total 94739701 7088627273 74 100.00%
Figure 3: Second Generation Firmware - RAID Level. File: stateCaptureData.txt
Virtual Disk Unit 0 Configuration
Volume Type: 13+1 RAID 5
User Label: MyRAID5_1
Block Size: 512 bytes
Large IO: 4096 blocks
Segment Size: 256 blocks
Stripe Size: 3328 blocks
...
IO Statistics:
small small large large cache
reads writes reads writes total hits
requests 2028332119 147699066 0 0 2176031185 1289775370
blocks 3091968111 2518067526 0 0 1315068341 4019884678
avg blocks 4 17 0 0 0 3
IO pct. 93.21% 6.78% 0.00% 0.00% 0.00% 59.27%
IOs stripes /IO clusters /IO
reads 2028332119 2034477363 1.00 2107869128 1.03
writes 147699066 148449472 1.00 157404718 1.06
write Full Partial RMW No Parity RMW2 FSWT
algorithms 1105611 12598366 32120072 0 0 0
Figure 2: First Generation Firmware - RAID Level. File: stateCaptureData.txt

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page16
to how tightly the variance of sequential or random data access is contained
within the volume. This can be purely random across an entire virtual disk, or
random within some bounds, such as a large file stored within a virtual disk
compared to large non-contiguous bursts of sequential data access randomly
distributed within some bounds. Each of these is a different I/O pattern, and has
a discrete case to be applied when tuning the storage.
The data from the stateCaptureData.txt file can be helpful in determining these
characteristics. Sequential read percentage can be determined based on the
percentage of total cache hits. If the cache hits and read percentage are high,
then first assume the I/O pattern tends to be more sequential I/O. However,
since cache hits are not broken out statistically by read and write, some variable
experimentation may have to be performed with the representative data set if the
pattern is unknown. For single threaded I/O host streams, this behavior can be
confirmed by comparing the magnitude of reads to read pre-fetch statistics.
In cases where many sequential read operations are expected, enabling read
pre-fetch in cache is recommended. If the cache hits percentage is low, the
application tends to be more random and read-ahead should be disabled. Mid-
range percentages possibly indicate bursts of sequential I/O but do not
necessarily denote their affiliation to read or write I/O. Again, testing with read-
ahead on/off would be required.
In the second generation firmware, the segment, stripe, and pre-fetch statistics
have been reorganized as seen in Figure 4 from the lower half of Figure 2.
4.7.4 Stripe Size
For the best performance, stripe size should always be larger than the maximum
I/O size performed by the host. As identified previously, stripes should be sized
as even powers of two. The average block size can be identified from the
collected data. Additionally, I/Os over 2MiB are considered large and broken out
separately from smaller I/Os in the statistics. While all RAID level’s benefit from
careful tuning of stripe and segment size, RAID 5 and 6 with their parity
calculations are the most dependant.
*** Performance stats ***
Cluster Reads Cluster Writes Stripe Reads
6252626 3015009 5334257
Stripe Writes Cache Hits Cache Hit Blks
2040493 4685032 737770040
RPA Requests RPA Width RPA Depth
982036 3932113 418860162
Full Writes Partial Writes RMW Writes
653386 29 328612
No Parity Writes Fast Writes Full Stripe WT
0 0 0
Figure 4: Second Generation Firmware - Performance
Statistics Broken Out. File: stateCaptureData.txt

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page17
In the first generation of firmware (see Figure 5), this can be determined from the
‘Avg. Blocks’ row which represents the average I/O block size encountered. In
Generation One, the ‘Large IO’ field denotes a 4096 block or 2MiB size with zero
registered Large Reads or Writes during the sample period. Any single host
received I/O greater than the Large I/O size is broken down into chunks less than
or equal to the stated Large I/O value. It is extremely rare for a host to send such
large I/Os.
With second generation firmware (see Figure 6), the ‘Avg. Blks’ column
represents the average the I/O block size encountered. In Figure 6, the
‘LargeIoSize’ field denotes a 2MiB size with zero registered Large Reads or
Writes during the sample period.
Volume 0 Attributes:
Volume Type: RAIDVolume
User Label: MyRAID10_One
...
BlockSize: 512 bytes
LargeIoSize: 4096 blocks
...
Perf. Stats: Requests Blocks Avg. Blks IO Percent
Reads 67456452 5943724625 88 71.20%
Writes 27283249 1144902648 41 28.80%
Large Reads 0 0 0 0.00%
Large Writes 0 0 0 0.00%
Total 94739701 7088627273 74 100.00 %
Virtual Disk Unit 0 Configuration
Volume Type: 13+1 RAID 5
User Label: MyRAID5_1
Block Size: 512 bytes
Large IO: 4096 blocks
Segment Size: 256 blocks
Stripe Size: 3328 blocks
...
IO Statistics:
small small large large cache
reads writes reads writes total hits
requests 2028332119 147699066 0 0 2176031185 1289775370
blocks 3091968111 2518067526 0 0 1315068341 4019884678
avg blocks 4 17 0 0 0 3
IO pct. 93.21% 6.78% 0.00% 0.00% 0.00% 59.27%
Figure 6: Second Generation Firmware - Volume Attributes Performance Stats; taken
from RAID 1 volume. File: stateCaptureData.txt
Figure 5: First Generation Firmware - Stripe Size. File: stateCaptureData.txt

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page18
In addition, the stateCaptureData.txt file provides a more granular method for
determining the distribution of I/O within stripes and segments. In Figure 7 and
Figure 8, item 1 is the number of full stripes read and written, and item 2
indicates the number of full clusters or segments read or written. The value of
stripes per I/O request within reads and writes is also valuable for determining if
stripe or segment settings are optimal for the tested data access pattern. The
second generation firmware does not specifically break out the per-I/O ratio in
data output as the first generation firmware did, however it can still be manually
calculated simply by dividing the value in item 1 or 2 by the appropriate I/O
request value from item 3 indicated in Figure 8.
In most cases, best performance is achieved with segment and stripe per-I/O
ratios as close to 1.00. Traditionally, when tuning for maximal I/Os per second, if
the segments per-I/O ratio is high, the current segment size may be too low for
the application. Similarly, when tuning for highest possible data transfer rate, the
stripe per-I/O ratio should ideally be 1.00 or even multiples. If this value is high,
increasing the number of physical disks and/or the segment size can improve
performance.
Virtual Disk Unit 0 Configuration
Volume Type: 13+1 RAID 5
User Label: MyRAID5_1
Block Size: 512 bytes
Large IO: 4096 blocks
Segment Size: 256 blocks
Stripe Size: 3328 blocks
...
IO Statistics:
small small large large cache
reads writes reads writes total hits
requests 2028332119 147699066 0 0 2176031185 1289775370
blocks 3091968111 2518067526 0 0 1315068341 4019884678
avg blocks 4 17 0 0 0 3
IO pct. 93.21% 6.78% 0.00% 0.00% 0.00% 59.27%
IOs stripes /IO clusters /IO
reads 2028332119 2034477363 1.00 2107869128 1.03
writes 147699066 148449472 1.00 157404718 1.06
write Full Partial RMW No Parity RMW2 FSWT
algorithms 1105611 12598366 32120072 0 0 0
Figure 7: First Generation Firmware– Stripe distribution. File: stateCaptureData.txt
1.
2.

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page19
4.7.5 Write Algorithm Data
It is important to understand the effect determining the most suitable RAID level
can be a daunting task. Understanding the effect of which write algorithm is in
use is an important part of RAID Level balance. The possible choices in first
generation firmware, as seen in Figure 9, are Full, Partial, RMW, RMW2, and the
Full Stripe Write-Through. RMW2 was folded into RMW statistics in the second
generation firmware (see Figure 10).
The Full algorithm takes an entire stripe of data and dumps it to the disk,
depending on the RAID level of choice, P or P and Q will get calculated at this
point. This is the most efficient type of write that can be performed, and the
design of a disk group should be around maximizing full writes.
Partial writes are when less than a full stripe of data non-aligned with segment
boundaries are modified and written. In RAID level 5 and 6 this is more complex,
as parity data must be recalculated for the whole stripe. Partial writes are a
worst-case algorithm and should be minimized. Larger quantities of partial writes
than full writes can indicate use of an inappropriate segment size.
Volume 0 Attributes:
Volume Type: RAIDVolume
User Label: MyRAID10_One
...
BlockSize: 512 bytes
LargeIoSize: 4096 blocks
...
Perf. Stats: Requests Blocks Avg. Blks IO Percent
Reads 67456452 5943724625 88 71.20%
Writes 27283249 1144902648 41 28.80%
Large Reads 0 0 0 0.00%
Large Writes 0 0 0 0.00%
Total 94739701 7088627273 74 100.33%
...
*** Performance stats ***
Cluster Reads Cluster Writes Stripe Reads
6252626 3015009 5334257
Stripe Writes Cache Hits Cache Hit Blks
2040493 4685032 737770040
RPA Requests RPA Width RPA Depth
982036 3932113 418860162
Full Writes Partial Writes RMW Writes
653386 29 328612
No Parity Writes Fast Writes Full Stripe WT
0 0 0
1.
2.
2.
3.
Figure 8: Second Generation Firmware– Stripe distribution. File:
stateCaptureData.txt

Dell™ PowerVault MD3000 and MD3000i Array Tuning Best Practices
December2008–RevisionA01 Page20
RMW, or Read-Modify-Write, is the second-best write algorithm available for
RAID 5 and 6. A RMW occurs when a quantity of bits, smaller or equal to an
individual segment are modified. This constitutes a two-read operation in RAID 5
and a three-read operation in RAID 6, with one of the segments being modified,
and the parity drive(s) are read in. Then parity for the affected region is
recalculated and then the data and parity in the stripe are re-written to disk. In
small transactional processing, an extremely large number of RMWs should be
expected. These RMW writes can cause a significant loss of performance;
however this impact can be lessened by proper tuning of virtual disk stripe size. .
RMW2 is used to differentiate Write-to-cache RMWs and Write-Through RMWs,
with RMW2 being the latter, these statistics were consolidated in the second
generation firmware. RMW2’s also specifically happen when cache is disabled
by force, failed mirroring controller (if policy is active) or failed cache battery.
Additionally, the second generation firmware discretely tracks Full Stripe Write-
Through conditions, and both generations track data on Number of Parity stripes
re-calculated.
Figure 9: First Generation Firmware - Write Algorithms
Virtual Disk Unit 0 Configuration
Volume Type: 13+1 RAID 5
User Label: MyRAID5_1
Block Size: 512 bytes
Large IO: 4096 blocks
Segment Size: 256 blocks
Stripe Size: 3328 blocks
...
IO Statistics:
small small large large cache
reads writes reads writes total hits
requests 2028332119 147699066 0 0 2176031185 1289775370
blocks 3091968111 2518067526 0 0 1315068341 4019884678
avg blocks 4 17 0 0 0 3
IO pct. 93.21% 6.78% 0.00% 0.00% 0.00% 59.27%
IOs stripes /IO clusters /IO
reads 2028332119 2034477363 1.00 2107869128 1.03
writes 147699066 148449472 1.00 157404718 1.06
write Full Partial RMW No Parity RMW2 FSWT
algorithms 1105611 12598366 32120072 0 0 0