Compaq 108164-003 - ProLiant - 800 Quick reference guide

WHITEPAPER
Help us improve our technical communication. Let us know what you think about the
technical information in this document. Your feedback is valuable and will help us structure
future communications. Please send your comments to: [email protected]
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ECG025.0997
[September 1997]
Prepared By
Microsoft Windows NT
Integration Team
Compaq Computer
Corporation
CONTENTS
Disk Subsystem
Overview
.....................
3
Disk-Related
Performance
Characteristics
............
4
Like Drive
Scalability
.................
17
Like Capacity
Scalability
.................
19
Disk Controller
Scalability
.................
24
Performance
Measurement Tools
...
27
Preventing Data
Loss while
Maintaining
Performance
..............
28
Disk Subsystem
Summary of
Findings
....................
29
Disk Subsystem Performance and
Scalability
In today's networking environments, the disk subsystem is a key element in determining
overall system performance. The goal of this paper is to provide informative test results
and performance-related information for various disk subsystems, to assist systems
engineers and network administrators in making decisions on disk subsystem installation,
optimization, and configuration.
This white paper also provides information on using Fault Tolerance to prevent data loss,
while maintaining system performance. Finally, this paper provides a section discussing
the advantages and disadvantages of RAID technology.

WHITE PAPER (cont.)
2
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
NOTICE
The information in this publication is subject to change without notice.
COMPAQ COMPUTER CORPORATION SHALL NOT BE LIABLE FOR TECHNICAL
OR EDITORIAL ERRORS OR OMISSIONS CONTAINED HEREIN, NOR FOR
INCIDENTAL OR CONSEQUENTIAL DAMAGES RESULTING FROM THE
FURNISHING, PERFORMANCE, OR USE OF THIS MATERIAL.
This publication does not constitute an endorsement of the product or products that were
tested. The configuration or configurations tested or described may or may not be the
only available solution. This test is not a determination of product quality or correctness,
nor does it ensure compliance with any federal, state or local requirements. Compaq
does not warrant products other than its own strictly as stated in Compaq product
warranties.
Product names mentioned herein may be trademarks and/or registered trademarks of
their respective companies.
Compaq, Contura, Deskpro, Fastart, Compaq Insight Manager, LTE, PageMarq,
Systempro, Systempro/LT, ProLiant, TwinTray, ROMPaq, LicensePaq, QVision, SLT,
ProLinea, SmartStart, NetFlex, DirectPlus, QuickFind, RemotePaq, BackPaq, TechPaq,
SpeedPaq, QuickBack, PaqFax, Presario, SilentCool, CompaqCare (design), Aero,
SmartStation, MiniStation, and PaqRap, registered United States Patent and Trademark
Office.
Netelligent, Armada, Cruiser, Concerto, QuickChoice, ProSignia, Systempro/XL, Net1,
LTE Elite, Vocalyst, PageMate, SoftPaq, FirstPaq, SolutionPaq, EasyPoint, EZ Help,
MaxLight, MultiLock, QuickBlank, QuickLock, UltraView, Innovate logo, Wonder Tools
logo in black/white and color, and Compaq PC Card Solution logo are trademarks and/or
service marks of Compaq Computer Corporation.
Other product names mentioned herein may be trademarks and/or registered trademarks
of their respective companies.
Copyright ©1997 Compaq Computer Corporation. All rights reserved. Printed in the
U.S.A.
Microsoft, Windows, Windows NT, Windows NT Server and Workstation, Microsoft SQL
Server for Windows NT are trademarks and/or registered trademarks of Microsoft
Corporation.
Disk Subsystem Performance and Scalability
First Edition (September 1997)
Document Number: ECG025.0997

WHITE PAPER (cont.)
3
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
DISK SUBSYSTEM OVERVIEW
Key components of the disk subsystem can play a major part of overall system
performance. Identifying potential bottlenecks within your disk subsystem is crucial. In
this white paper we identify and discuss in detail disk-related performance characteristics
that can help you understand how latency, average seek time, transfer rates and file
system or disk controller caching can affect your disk subsystem performance. Once we
discuss all of the disk measurement terms, we use those definitions to address
performance issues in each scalability section of this document. The different scalability
sections discussed are as follows:
•Like Drive (similar hard drive scalability)
•Like Capacity (similar drive capacity scalability)
•Disk Controller (multiple controller scalability)
This document provides disk subsystem recommendations, based on testing in the
Integration Test Lab of hardware and software products from Compaq and other vendors.
The test environment that Compaq selected might not be the same as your environment.
Because each environment has different and unique characteristics, our results might be
different than the results you obtain in your test environment.
Test Environment
The following table describes the test environment used for the disk subsystem
performance testing. This table displays both the one and two controller test
configurations that were used in the Compaq ProLiant 5000 during testing.
Table 1:
Disk Subsystem Testing Environment
Environment Equipment Used
Server Hardware Platform Compaq ProLiant 5000
Memory 128 MB
Processors (4) P6/200 MHz 512k secondary cache
Network Interface Controllers (2) Dual 10/100TX PCI UTP Controller (4 network segments)
Disk Controllers (1 or 2) SMART-2/P Controllers
Disk Drives 2, 4, or 9 GB Fast-Wide SCSI-2 drives
Number of Drives up to fourteen 2.1, 4.3, or 9.1 GB Fast-Wide SCSI-2 drives
Boot Device (1) Fast-Wide SCSI-2 drive off the Embedded C875 controller

WHITE PAPER (cont.)
4
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Table 1: (cont.)
Disk Subsystem Testing Environment
Environment Equipment Used
Server Software Configuration Microsoft Windows NT Server version 4.0
Service Pack 2
Compaq Support Software Diskette 1.20A
Client Configuration Compaq Deskpro 575
Netelligent 10/100 TX PCI UTP Controller
and MS-DOS
NetBench 5.0 Test Configuration Disk Mix
Work Space 15 MB
Ramp Up Time 10 seconds
Ramp Down Time 10 seconds
Test Duration 120 seconds
Delay Time 0 seconds
Think Time 0 seconds
DISK-RELATED PERFORMANCE CHARACTERISTICS
Before beginning our discussion on disk subsystem performance, Table 2 lists the general
terms used in the industry to describe the performance characteristics of disk
performance. These general terms describe characteristics that can impact system
performance, so it is important to understand the meaning of each term and how it could
affect your system.
Table 2:
Disk Performance Measurement Terms
Terms Description
Seek Time The time it takes for the disk head to move across the disk to find a
particular track on a disk.
Average Seek Time The average length of time required for the disk head to move to the
track that holds the data you want. This average length of time will
generally be the time it takes to seek half way across the disk.

WHITE PAPER (cont.)
5
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Table 2: (cont.)
Disk Performance Measurement Terms
Terms Description
Latency The time required for the disk to spin one complete revolution.
Average Latency The time required for the disk to spin half a revolution.
Average Access Time The average length of time it takes the disk to seek to the required
track plus the amount of time it takes for the disk to spin the data
under the head. Average Access Time equals Average Seek Time
plus Latency.
Transfer Rate The speed at which the bits are being transferred through an
interface from the disk to the computer.
Concurrency The number of I/O requests that can be processed simultaneously.
RPM (Revolutions Per Minute) The measurement of the rotational speed of a disk drive on a per
minute basis.
Table 2 lists the definitions of disk-related performance characteristics. Let’s now use
those definitions in the next several sections to address how adding drives to your system
can affect performance.
Seek Time and Average Seek Time
Seek time describes the time it takes for the disk head to move across the disk to find
data on another track. The track of data you want could be adjacent to your current track
or it could be the last track on the disk. Average seek time, however, is the average
amount of time it would take the disk head to move to the track that holds the data.
Generally, this average length of time will be the same amount of time it takes to seek half
way across the disk and is usually given in milliseconds.

WHITE PAPER (cont.)
6
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
One method to decrease seek time is to distribute data across multiple drives. For
instance, the initial configuration in Figure 1 shows a single disk containing data. The
new configuration reflects the data being striped across multiple disks. This method
reduces seek time because the data is spread evenly across two drives instead of one,
thus the disk head has less distance to travel. Furthermore, this method increases data
capacity because the two disks provide twice the space to store data.
Data
Seek Time
Initial
Configuration
New
Configuration
Unused space Unused space
Data
Seek Time Seek Time
Unused space
Data
Figure 1: Average Seek Time and Capacity
You can use this same concept and apply it to many different configurations. For
example, if you currently have a two-disk configuration but you want to decrease the
average seek time yet increase the disk capacity, you can configure a striped set of disks
using four disks instead of two. This concept applies to any configuration (odd or even
number of disks) as long as you are adding more drives to your stripe set.

WHITE PAPER (cont.)
7
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Average Latency
Manufacturers have built and continue to build hard disks that spin at designated rates. In
the early years of the personal computer (PC) industry, hard disks on the market could
spin at approximately 3600 RPMs. As the market demand for better system performance
increased, disk manufacturers responded by supplying faster spin rates for hard disks.
By producing faster spinning disks, manufacturers reduced the amount of overall access
time. Average latency directly correlates to the spin rate of the disk drive because it is, as
defined earlier in Table 2, the time required for the disk to spin half a revolution.
Therefore, this direct relationship in improving hard disk spin rates can contribute to better
system performance by reducing the average latency on a disk.
Manufacturers understand the need for better system performance and continue to
provide new and improved hard disks. With today’s hard disks spinning at 7200
revolutions per minute (RPMs) and the hard disks of tomorrow spinning at the rate of
10,000 RPMs, we can see that manufactures continue to address the issue of faster
performance. Table 3 provides a brief history on hard disks listing spin rates, disk
capacities available and approximate dates the disks were available to the market.
Table 3:
Hard Disk History
Disk Spin Rate Disk Capacity Approximate Date Used
3600 RPMs Up to 500 MB 1983 – 1991
4500 RPMs 500 MB – 4.3 GB 1991 – Present
5400 RPMs 500 MB – 6 GB 1992 – Present
7200 RPMs 1 GB – 9.1 GB 1993 - Present
10,000 RPMs 4.3 GB and 9.1 GB 1997 - Present
Now that we discussed the direct relationship between disk spin rates and system
performance, let’s examine how drive scaling can affect latency. In Figure 2 - Example 1,
the initial configuration shows the disk has to spin halfway around before the disk head
can start to retrieve data from sector 5. In the new configuration, the disk has to spin half
the distance than before to retrieve the same data. Thus, the latency time has been cut
in half.

WHITE PAPER (cont.)
8
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
However, be aware that the average latency time might not always decrease when adding
more drives to your system. For example, in Figure 2 - Example 2, the new configuration
shows that the amount of time it takes to retrieve the data from sector B is actually longer
than the initial configuration. The reason for this is that the disk has to spin half way
around to read sector B. In the initial configuration the disk only had to spin one-eighth a
revolution to read the identical data. But, keep in mind that the initial configuration for
Example 2 required both seek time and latency time.
Initial
Configuration
New
Configuration
Example 2
1
A
2B
3C
4D
5
E
6
F
G
H7
8
5
3
1
7A
C
E
G
6
4
2
8B
D
F
H
Disk Head
Disk Head Disk Head
Example 1
Example 1
Example 2
Figure 2: Average Latency
Overall, these examples show us that in some configurations, as shown in our first
example, drive scaling would be a definite performance advantage. However, in other
configurations it is not clear if you receive a performance gain because of the components
involved, such as the combination of average seek time and average latency time used in
Figure 2 – Example 2.
When you combine these terms (average seek time and average latency time), you define
another disk measurement called average access time, which is discussed in the
upcoming section. From the information provided in this section, we know seek time plus
latency (or average access time) is a key in determining if performance is truly enhanced
in your system.
Note: Remember that the disks
used in Figure 2 are identical in
size and RAID configuration.

WHITE PAPER (cont.)
9
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Average Access Time
Average access time is simply described as average seek time plus latency. What this
equates to is the amount of time the disk has to seek to find the data plus the time it takes
for the disk to spin under the head. For example, Figure 3 contains a disk with two tracks
of data on it. Track 1 contains data sectors 1 – 8. Track 2 contains data sectors A – H.
Thus, in our example, the disk head has to move (or seek) from the current position (track
1, sector 1) to the track you want to read (track 2, sector C).
For the purpose of our illustration, Figure 3 displays the disk head performing these
functions separately. However, in reality the disk drive performs both seek and latency
functions simultaneously.
Unused space
Data
1
A
2B
3C
4
D
5
E
6
F
G
H7
8
Disk Head
Latency
Seek Time
Average Seek Time + Latency = Average Access Time
Figure 3: Average Access Time
Transfer Rates
A disk subsystem is made up of multiple hardware components that communicate by
transferring data to and from the disk(s) to a computer. The main parts of a disk
subsystem are as follows:
•Hard Disks
•SCSI Channel
•Disk Controller
•I/O Bus
•File System and Disk Controller Caching

WHITE PAPER (cont.)
10
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
In order to share information, all of the disk subsystem components have to communicate
with each other, as shown in Figure 4. The disk subsystem components communicate
with each other using hardware interfaces such as Small Computer System Interface
(SCSI) channels and Peripheral Component Interconnect (PCI) buses. These
communication highways, called channels and/or buses, communicate at different rates of
speed known as transfer rates.
Each of the disk subsystem components transfer data at different rates. It is important to
understand the different transfer rates of each component because this information helps
you identify potential performance bottlenecks within your system. For example, Figure 4
shows hard disks transferring data to the SCSI channel (bus), which transfers the
information to the disk controller, which then passes the data to the Host Bus and then on
to the server. If one hard disk transfers at 5 MB/s, the SCSI channel transfers at 40 MB/s,
the disk controller transfers at 40 MB/s and the Host Bus transfers at 540 MB/s, it is
obvious that the hard disk is the bottleneck. Therefore, by knowing the transfer rate of
each subsystem device, potential bottlenecks can be easily identified and corrected.
The key to improving system performance is focusing on how to maximize data
throughput by minimizing the amount of time the disk subsystem has to wait to receive or
send data. In the upcoming sections, we discuss how to identify performance bottlenecks
and where they could possibly occur in your disk subsystem.
PCI Bus transfers
to/from the Host Bus
More Drives
Controller transfers to/from disks via SCSI channel
Disk drive average
sustained transfer
rate
SCSI Bus
transfer rate
PCI Bus
transfer rate
1
1
3
4
2
2
Host Bus (Memory)
transfer rate
3
Host Bus transfers to/from
ProLiant 5000
4
Controller transfers
to/from PCI Bus
Dual Inline Memory
Module (DIMM)
Figure 4: Disk subsystem components transferring data.
Remember that the slowest disk
subsystem component determines
the overall throughput of the
system.

WHITE PAPER (cont.)
11
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Disk Transfer Rates
Hardware manufacturers calculate and define disk transfer rates as being the theoretical
threshold for transferring data from the disk to the computer. For example, if you were to
place one drive with an average transfer rate of 5 MB/s (see 1in Figure 4) in a system,
theoretically it would take four disks to saturate a SCSI channel with a transfer rate of
20 MB/s (see 2in Figure 4).
If you were to saturate the disk subsystem by adding drives, concurrency would increase
because the system is able to process more I/O requests. Thus, increasing overall
throughput, which improves system performance. A detailed discussion on concurrency
is provided later in this document.
It is important to note the difference between average disk sustained transfer rate and the
transfer rate of the SCSI bus. In the disk transfer example above, the average disk
sustained transfer rate refers to the disk transferring data at 5 MB/s. This transfer rate is
a completely separate performance rating than the SCSI bus transfer rate, which in our
example is 20 MB/s. Disk drives have a special interface used to communicate with the
SCSI bus. This interface, defined in disk drive characteristic specification documents,
identifies the type of controller the drive supports not the transfer rate of the disk. For
example, if you are using a Wide-Ultra drive you know that this drive supports the
Wide-Ultra SCSI Controller, which transfers at 40 MB/s but the average disk sustained
transfer rate for the drive might be 5 MB/s.
The earlier example listed above provides a simple illustration of a disk transferring data
at 5 MB/s. However, some hard disks being manufactured today transfer data faster than
the disk in the example. Table 4 lists the transfer rate specifications for all of the hard
disk drives used during lab testing.
Table 4:
Hard Disk Transfer Rates
Disk Capacity Defined Transfer Rate Average Sustained TransferRate
2.1 GB Up to 40 MB/s 4 MB/s
4.3 GB Up to 40 MB/s 5 MB/s
9.1 GB Up to 40 MB/s 7 MB/s
Note: The actual transfer rates
listed in Table 4 depend on the
type of I/O being performed in
the system.

WHITE PAPER (cont.)
12
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
SCSI Channel Transfer Rates
The disk controllers being used today can transfer data up to 40 MB/s to and from the
hard disk to the disk controller by way of the SCSI bus. However, if your disk drive can
sustain a transfer rate of only 5 MB/s, the SCSI bus is going to be idle 87.5% of the time.
In this example, the disk drive is the bottleneck because it transfers data slower than the
SCSI bus.
The key to improving system performance is to maximize data throughput by minimizing
the time the disk subsystem has to wait to send or receive data. If the cumulative
sustained transfer rate of the drives is less than the transfer rate of the SCSI channel,
there is a significant chance that the drives will limit the throughput. Alleviate the disk
bottleneck by adding additional drives to the system. For maximum performance, the
total disk transfer rate should be equal to or greater than the SCSI channel transfer rate.
For example, if the SCSI channel transfer rate is 40 MB/s (Wide-Ultra SCSI), add six
9.1 GB drives (6 x 7 MB/s = 42 MB/s) to reach a sustained transfer rate equal or greater
than the SCSI channel.
Disk Controller Transfer Rates
Disk controllers are continuously being upgraded to support wider data paths and faster
transfer rates. Currently, Compaq supports three industry standard SCSI interfaces on
their disk controllers, as shown in Table 5.
Table 5:
Compaq Disk Controllers
Controller Name Description
Compaq Fast-SCSI-2 SCSI interface that uses an 8-bit data path with transfer rates up
to 10 MB/s
Compaq Fast-Wide SCSI-2 SCSI interface that uses a 16-bit data path with transfer rates up
to 20 MB/s
Compaq Wide-Ultra SCSI SCSI interface that uses a 16-bit data path with transfer rates up
to 40 MB/s
Disk controllers can be a common cause of disk subsystem bottlenecks. For example, if
a disk subsystem contains a Compaq Wide-Ultra SCSI Controller transferring data up to
40 MB/s, ideally it would take three controllers to saturate the PCI Bus, which transfers
data at the rate of 133 MB/s. Again, similar to the disk transfer rate example discussed
earlier, concurrency would increase once you begin to add more controllers to the disk
subsystem. The additional controllers enable the system to process more I/O requests,
thus improving overall system performance.
SCSI Bus Idle Time can be
calculated as follows:
[sustained transfer rate] x
[number of drives]
÷
[SCSI bus
transfer rate]= amount of time
the SCSI bus is busy. Subtracting
this result from 1 provides the
SCSI bus idle time. For example,
[5 MB/s] x [1 drive]
÷
[40 MB/s]
= 12.5%. Subtracting this value
from 1 equals 87.5%.

WHITE PAPER (cont.)
13
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
I/O Bus Transfer Rates
The I/O Bus consists of one or more of the following: Peripheral Component Interconnect
(PCI), Dual Peer PCI, Extended Industry Standard Architecture (EISA) or Industry
Standard Architecture (ISA). Table 6 defines these I/O bus types and their transfer rates.
Table 6:
I/O Bus Transfer Rates
I/O Bus Type Definition and Transfer Rate
Peripheral Component Interconnect (PCI) A system I/O bus architecture specification that supports 32-bit
bus-mastered data. Designed to support plug-and-play
configuration of optional peripherals. Transfers at a maximum
rate of 133 MB/s.
Dual Peer PCI
(Supported on the ProLiant 5000, 6000,
6500 and 7000)
A system I/O bus architecture specification that supports 32-bit
bus-mastered data. Designed to support plug-and-play
configuration of optional peripherals. Each controller transfers at
a maximum rate of 133 MB/s, with a combined total throughput
of 266 MB/s.
Extended Industry Standard Architecture
(EISA) A system I/O bus architecture specification that supports 8-, 16-
and 32-bit data throughput paths. Supports bus-mastering on 16-
and 32-bit buses. Transfers at a maximum rate of 33 MB/s.
Industry Standard Architecture (ISA) A system I/O bus architecture specification that supports 8- and
16-bit data throughput paths. Supports bus-mastering on 16-bit
buses. Transfers at a maximum rate of 8 MB/s.
The theoretical threshold for the PCI Bus has a transfer rate of 133 MB/s. Because the
PCI Bus can transfer data so quickly, it is the second least (file system cache being the
first) likely of all of the disk subsystem components to be a performance bottleneck. To
illustrate the point, it would take a minimum of three Compaq Wide-Ultra SCSI Controllers
running at their maximum sustained transfer rate of 40 MB/s each to maintain throughput
on the PCI Bus. Even running this configuration (3 x 40 MB/s = 120 MB/s) does not
completely saturate the PCI Bus, having the capability of transferring at a rate of
133 MB/s.
File System and Disk Controller Caching Transfer Rates
File system and disk controller caching plays a fundamental role in system performance.
Accessing data in memory, also known as Random Access Memory (RAM), is extremely
fast (refer to Table 7 for transfer rates). Accessing data on the disk is a relatively slow
process. If, in theory, we could avoid disk access by requesting and retrieving data from
memory or “cache”, system performance would improve dramatically.
Note: The File System Cache
data is stored in memory.
Accesses to this data takes place
over the Host Bus.

WHITE PAPER (cont.)
14
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
For instance, let’s say you request data stored on your disk drive (refer to Figure 4 for
reference). The system first tries to complete the READ request by retrieving the data
from the file system cache (memory). If it is not there, the system has to retrieve the data
from the hard disk. Figure 5 shows the communications that take place to retrieve data
from the disk.
Operating system
checks to see if the
requested data is in
memory.
No
Yes
Operating system sends the
request through the I/O bus
(i.e., PCI bus) to the disk controller.
Yes
Disk controller keeps a copy of
the data in its cache and passes
the data through the I/O bus
(i.e., PCI bus).
Disk controller sends the request
across the SCSI bus to the
physical disk drives.
Drive head(s) seek to the track
where the data is located and reads
the data.
Drive transfers the data across the
SCSI bus to the disk controller.
Operating system receives the data, passes it to
the requesting application and keeps a copy of
the request in its cache.
Disk controller
checks its cache (i.e.,
Array Accelerator) for
the data.
No
Figure 5: Retrieving data from the hard disk.
Now let’s examine the same scenario if the requested information were located in file
system cache (refer to Figure 5 if necessary). The operating system checks to see if the
requested data is in memory (i.e., File system cache). Operating system passes the
requested information to the application.
As you can conclude from the flowchart example, the more information stored in memory
the faster the system can access the requested data. Thus, if you are retrieving data from
memory, the speed of the Host Bus will influence system performance.

WHITE PAPER (cont.)
15
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Table 7 lists the Host bus transfer rates for the following Compaq servers:
Table 7:
Host Bus (Memory) Transfer Rates
Server Name Transfer Rate
ProLiant 5000, 6000, 6500 and 7000 540 MB/s
ProLiant 1500, 2500 and 4500 267 MB/s
The last example discussed how READ performance is increased. Now let’s discuss how
WRITE performance is enhanced on a system by taking advantage of posted writes.
Posted writes take place when file system or disk controller caching temporarily holds one
or more blocks of data in memory until the hard disk is not busy. The system then
combines or “coalesces” the blocks of data into larger blocks and writes them to the hard
disk. This results in fewer and larger sequential I/Os. For example, a network server is
used to store data. This server is responsible for completing hundreds of client requests.
If the server happened to be busy when data was being saved, the server’s file system
cache tells the application that the data has been saved so that the application can
continue immediately without having to wait for the disk I/O to complete.
Coalescing is also commonly referred to as “Elevator seeking.” This coined phrase
became popular because it provides the perfect analogy for describing coalescing. For
instance, an elevator picks up and drops off passengers at their requested stop in the
most efficient manner possible. If you were on level 6, the elevator on level 2, and other
passengers on levels 1 and 7, the elevator would first stop on level 1 to pick up the
passenger going up. Next, the elevator would stop on level 6, then 7 and then take
everyone to level 9, their destination. The elevator would not perform all of the requests
individually, instead it reorders then completes those requests in a more efficient manner.
This same analogy applies to coalescing when writing data to different sectors on a disk.
As an example, Joe saves or “writes” data B to the hard disk, then he saves data A to the
same disk. And finally, he saves data C as well. Instead of completing 3 separate I/Os
for B, A, then C, the system reorders the write requests to reflect data ABC then performs
a single sequential I/O to the hard disk, thus improving disk performance.
Note: Compaq uses coalescing
algorithms to optimize disk
performance.

WHITE PAPER (cont.)
16
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Concurrency
Concurrency is the process of eliminating the wait time involved to retrieve and return
requested data. It takes place when multiple slow devices (e.g., disk drives) place I/O
requests on a single faster device (e.g., SCSI bus). As shown in Figure 6, a request for
data comes across the SCSI Bus asking the disk drive to retrieve some information. The
disk drive retrieves then sends the requested data back to the server via the SCSI bus.
The time it takes to complete this process seems to be acceptable at first glance until you
examine the amount of time the SCSI bus remains idle. This idle time shown in Figure 6
is the amount of time the SCSI bus is waiting for the disk drive to complete the request.
This valuable time could be used more efficiently in an environment taking advantage of
concurrency.
SCSI Bus
Idle
Time
Drive 1 1 1 4
2 3 3 4 4
32
2
134 2
Legend
Request for Data (SCSI Bus)
Retrieving Data (Disk Drive)
Sending Data (SCSI Bus)
Idle Time (SCSI Bus)
First I/O Request
Figure 6: I/O request timing diagram for a single drive configuration.

WHITE PAPER (cont.)
17
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Concurrency is very effective in a multi-drive environment because, while one drive is
retrieving data, another request can be coming across the SCSI bus as shown in
Figure 7. When using multiple drives, each drive can send data across the SCSI bus as
soon as it is available. As more drives are added to the system, the busier the SCSI bus
becomes. Eventually the SCSI bus becomes so busy that it yields no idle time.
SCSI Bus
Idle
Time
Drive 1
Drive 2
Drive 3
1 1 1
2 2 2
3 3 3
4 4 4
5 5 5
6 6 6
Idle Time (Disk Drive)
Legend
Request for Data (SCSI Bus)
Retrieving Data (Disk Drive)
Sending Data (SCSI Bus)
Idle Time (SCSI Bus)
First I/O Request
Figure 7: Concurrency taking place in a multiple drive configuration.
In conclusion, without concurrency (shown in Figure 6) the SCSI bus remains idle 60% of
the time. In contrast, when using concurrency (shown in Figure 7) the SCSI bus remains
busy 100% of the time and the subsystem is able to transfer more I/O requests in a
shorter period of time.
In the next few sections, we apply the knowledge learned earlier in this document to
analyze the test results for Like Drive, Like Capacity and Disk Controller Scaling.
LIKE DRIVE SCALABILITY
Hardware scalability is difficult to accomplish and to maintain. The right balance or
mixture is crucial for an effective disk subsystem. It is important to remember to balance
the current performance needs with future disk capacity and performance requirements.
For this reason, you need to choose the best performance configuration for the current
disk subsystems, and at the same time allow enough room in the configuration to fulfill

WHITE PAPER (cont.)
18
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
growth and increasing capacity requirements. For instance, choose the correct server
configuration for your environment, yet leave slots available for future disk controllers that
might be necessary to support future capacity requirements.
Like Drive Scaling
Like drive scaling is defined as comparing similar drives with the same RAID level and
“scaling” or adding more drives to your system so that you can measure cumulative disk
performance. Drive scaling can be summarized as the more drives you add to your
system, the better the performance. However, the question you need to ask yourself is
“When does the cost of adding more drives out weigh the performance gain?”.
To answer this question, Compaq tested controllers using the same RAID configuration
and added drives, then measured the system performance effects. Let’s now view those
results and understand the effects of drive scaling.
Like Drive Scalability Test Results
Our one controller testing, illustrated in Figure 8, revealed that when using 2GB drives
configured in a RAID 0 environment approximately a 50% performance increase was
achieved when the drives were doubled. For example, we doubled the number of drives
in the test configuration environment from 4 to 8 and experienced a 57% increase
in performance.
Like Drive Scaling (RAID 0 - 1 Controller)
-
2,000,000.00
4,000,000.00
6,000,000.00
8,000,000.00
10,000,000.00
12,000,000.00
14,000,000.00
16,000,000.00
18,000,000.00
20,000,000.00
4 8 16 20 24 28 32 36 40 44 48 52 56 57
Number of Clients
Server Throughput (Bytes/sec)
12 x 2GB
8 x 2GB
4 x 2GB
57%
25%
96%
Figure 8: Like Drive Scaling in a RAID 0 Environment.
In addition, we found when using 2GB drives in a RAID 5 configuration (4+1 vs. 8+1
drives), performance also increases over 50%. A performance increase was also
obtained when using 4GB drives on one controller in either a RAID 0 or 5 environment.
A discussion of RAID levels used
in this white paper can be found
in the “Preventing Data Loss
While Maintaining Performance”
section of this document.

WHITE PAPER (cont.)
19
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Lastly, our one controller testing shows that once we added another four drives to our test
environment (8 to 12 drives), the increase in performance was 25%, as shown in Figure 8.
Summary of Findings – Like Drive Scaling
Doubling the number of drives in our system, in either a RAID 0 or 5 environment,
increased performance by more than 50% when using 2 or 4GB drives. Also, by adding
four more drives to our environment (making a total of 12 drives) as shown in Figure 8, we
learned that our performance gain increased another 25%. However, keep in mind that
by doubling our drives in our test environment we improved performance, but we also
doubled the drive cost for this configuration. If performance is a concern, this is an
effective solution for your environment.
To help assist you in deciding what is right for your environment, Table 8 lists some
advantages and disadvantages of like drive scaling.
Table 8:
Like Drive Scaling Advantages and Disadvantages
Advantages Disadvantages
Eliminates bottlenecks becauseyou
minimize seek time. Drive cost increases each time you add more drives.
Increases I/O concurrency because you
have more drives processing disk requests. Using smaller size drives limits your maximum capacity per controller. For example, the Compaq SMART-2 Array Controller
supports up to 14 drives. By using fourteen 2GB drives, your data
capacity equals 28GB. By using fourteen 4GB drives, your data
increases to 56GB.
Increases cumulative transfer rate because
the more drives you add to your system the
more data can be transferred.
By adding more drives to your system you have more disks to
manage, thus increasing the probability of disk failure.
Idle time on the controller decreases
because cumulative disk performance
increases.
LIKE CAPACITY SCALABILITY
Like capacity scaling, unlike like drive scaling, is when you use similar or “like” drives and
scale them to determine if your environment needs multiple lower capacity drives or fewer
larger capacity drives. For example, if you need four Gigabytes of disk storage, what
should you purchase to meet your capacity requirements and provide the best system
performance? Should you buy one 4.3-Gigabyte drive or two 2.1-Gigabyte drives? If you
currently have hard drives in your environment that are not using the total storage
capacity of the disk, using like drives with a smaller capacity might be the right solution
for you.

WHITE PAPER (cont.)
20
ECG025.0997
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Like Capacity Scaling
Since like capacity scaling can affect your system, it is important to understand the impact
it might have on system performance. To be able to determine this information, we tested
like capacity scalability by maintaining the same total disk capacity for each test (8GB or
24GB) and added different quantities of drives to a single disk controller. The results of
these tests determine if system performance improves when using multiple lower capacity
drives or fewer larger capacity drives. In the next few sections of this white paper we
analyze these different configurations.
Like Capacity Test Results
Earlier we discussed concurrency and how the more spindles (disks) you have in your
system the better the performance would be because more I/O requests are being
concurrently processed. Overall our single disk controller like capacity tests provide
evidence that support our theory on concurrency. For example, if you need 8 Gigabytes
of storage capacity, our test show the benefits of using four 2GB disks instead of two 4GB
disks. The storage capacity is the same; however, the performance increase is 68% as
shown in Figure 9. We needed only one 9GB drive in our test to reach the eight Gigabyte
storage capacity requirement, so consequently concurrency could not take place and is
therefore not beneficial in this configuration.
Like Capacity Scaling (RAID 0 - 1 Controller)
0
2,000,000
4,000,000
6,000,000
8,000,000
10,000,000
12,000,000
14,000,000
16,000,000
4 8 16 20 24 28 32 36 40 44 48 52 56 57
Number of Clients
Server Throughput (Bytes/sec)
4 x 2GB
2 x 4GB
1 x 9GB
68%
Figure 9: Like Capacity Scaling in a RAID 0 Environment.