Understanding IBM eServer pSeries Performance and Sizing

Chapter 4. IBM RS/6000 and IBM pSeries products 115

Following is an approximation of free space required in various filesystems

and volume groups:

• The recommendation for rootvg is 2 GB, providing that /spdata is installed

in a separate volume group.

• /var requires at least 20 MB of free space. Most of the PSSP logs are

stored here. The actual disk space used by these logs depends entirely on

the types of problems that occur on your system and their frequency. The

/var filesystem should be monitored frequently to ensure that there is

always sufficient free disk space for new logs.

• /tmp requires at least 16 MB of free space.

• We suggest that you create a separate file system for /tftpboot. Each

lppsource level requires a minimum of 25 MB. You should also consider

future AIX installations. We suggest 25 MB x (number of AIX versions+1).

For example, if you are installing AIX 4.2, 4.2.1, and 4.3, then you will

need 25 MB x (3+1) = 100 MB.

• Downloading all of the AIX filesets to the lppsource directory requires

approximately 1.5 GB of disk space. Downloading only the minimal filesets

requires approximately 500 MB. Each mksysb image may vary between

100 MB and 700 MB. A simple example for AIX 4.3.2 is:

lppsource + mksysb_images + pssp_lpp_images + SPOT = total disk

space

lppssource = 500 MB

mksysb_images = 300 MB

pssp_lpp_images = 350 MB

SPOT = 200 MB

500 MB + 300 MB + 350 MB + 200 MB = 1.35 GB

You should have at least 1.35 GB of free disk space for the /spdata file

system. To be on the safe side, reserve 2 GB of disk space for /spdata. For

multiple AIX and PSSP version coexistence, this figure will increase

considerably due to the disk space taken up by different versions of

mksysb images, SPOT, AIX, and PSSP filesets.

4.2.6 Sizing and configuring an SP system

Since the initial introduction of the IBM RS/6000 SP systems, IBM has

continued to announce new node types with faster performance and better

price/performance ratios. Today, all the types of SP nodes are SMP nodes.

116 RS/6000 and IBM ^ pSeries Performance and Sizing

4.2.6.1 Node selection

When selecting nodes for an SP system, we can use the methodology shown

in Figure 45 as a general rule.

Figure 45. Node type selection

* The 332 MHz PowerPC node is supported but a new POWER3 node is

recommended.

1. Scientific/Technical Application

A scientific/technical application generally needs high floating-point

processor performance. The POWER3 chip has higher floating-point

processor performance than PowerPC 604e. The SPECfp_base_rate95 of

the POWER3 SMP 8way High Node is 1760, and that of the 332 MHz SMP

4way Wide/Thin Node is 364.

Also, if your application requires large memory, the high node is

recommended. Because the maximum memory size of the POWER3 SMP

High Node is 16 GB and that of the POWER3 Wide and Thin Node is 4

GB.

2. Commercial Application

If you use a commercial application, consider how much I/O is required.

The S80/S7A SP-Attached Node is recommended when very heavy I/O

Application

Type?

POWER3

Nodes

332 MHz

SMP

Wide

Nodes

S80

SP-Attached

Nodes

How much

Memory

needed?

Scientific/Technical

How much

I/O

Required?

Commercial

POWER3

High Nodes

POWER3

Wide/Thin

Nodes

Heavy

Medium/

Large

General

Large

Scale?

Ye s

No

Light

*

Start here

Chapter 4. IBM RS/6000 and IBM pSeries products 117

and very large scalability are required. If the I/O is heavy but not large

scale, you can chose POWER3 SMP Nodes. If the I/O is medium or light,

332 MHz SMP Wide Node is recommended. Commercial applications

generally need many PCI slots.

4.2.6.2 Parallel sizing factors

Sizing a scalable system for commercial applications is different from sizing a

non-scalable system. It depends on the type of commercial applications, and

the following factors should be considered. Usually there are two types of

commercial applications predominately used on a scalable system; Decision

Support System (DSS) and On-line Transaction Processing (OLTP).

The performance metrics for the system designed for OLTP applications are:

• Throughput

• Response time

• N-node scale-up

The performance metrics for the system designed for DSS applications are:

• Throughput

• Response time

• N-node scale-up

• N-node speed-up

• Database scale-up

Scale-up and speed-up

In the case of scale-up, twice as much hardware capacity typically should

perform twice the work in the same elapsed time. Scaling up a scalable

system can be done in two ways, as shown in Figure 46 on page 118. The

first case occurs when increasing the data and resources equally. In this

case, perfect linear scaling will produce the same elapsed time. In the second

case, the amount of data increases while the resources stay constant. In the

latter case, elapsed time should increase proportionally to the workload

increase.