Measurements; The Inetload Tool - Compaq 108164-003 - ProLiant - 800 White Paper

W P
HITE APER
(cont.)
. .
. .
. .
P
. . E R F O R M A N C E
. .
. .
The goal of capacity planning is to determine the types and configurations of computer systems,
. .
networks, and other resources that can support a specified number of users with a precise response
. .
. .
time. In order to do this with realistic resources and in a reasonable amount of time, it is
. .
necessary to simulate many users interacting with the System Under Test (SUT), in this case the
. .
. .
server being tested. The simulation tools used in this study were Microsoft's most recent versions
. .
of InetLoad and WCAT, as well as some specially developed Web pages using Microsoft ASP
. .
. .
facilities.
. .
. .
The most important performance measurement needed to aid in capacity planning is the
. .
maximum transaction rate for a specified load that a computer system can support. A system
. .
. .
reaches its maximum transaction rate when its CPU or CPUs are utilized 100% of the time. This
. .
metric is called the "platform transaction capacity", and it is highly dependent on the mix of
. .
. .
transactions used for its measurement. For example, a Web server that is tested transferring files
. .
that average 500 bytes to a simulated browser might have a platform transaction capacity of 2,000
. .
. .
operations per second. Another test of the same computer transferring an average file size of
. .
7,800 bytes, which is a much more realistic average size for Web pages and graphics, would show
. .
. .
a platform transaction capacity of 900.
. .
. .
There are two ways to measure platform transaction capacity. The first is to simulate many people
. .
using a system with a specified sequence of transactions, including time for a person to read or
. .
. .
think about the information returned by the system. InetLoad supports this approach. The second
. .
way to measure platform transaction capacity is to generate transactions as quickly as possible, in
. .
. .
a random sequence, but with a specified frequency. Both InetLoad and WCAT support this
. .
approach.
. .
. .
When measuring platform transaction capacity, it is imperative that the system's CPUs are the
. .
. .
limiting factor, not some other aspect of the environment. This clearly requires a controlled and
. .
well defined environment. Other limitations, or bottlenecks, can greatly affect measured platform
. .
. .
transaction capacity. Microsoft's perfmon performance monitoring tool, which comes with
. .
Windows NT Server 4.0, was used to help determine the existence and location of bottlenecks.
. .
. .
These limitations are discussed in the "Bottlenecks" section of this document.
. .
. .
. .

The InetLoad Tool

. .
. .
InetLoad supports testing multiple protocols, including the Lightweight Directory Access Protocol
. .
. .
(LDAP), which is needed for ILS testing, and the HyperText Transfer Protocol (HTTP), which is
. .
necessary for PS testing. It works by having worker threads execute scripts on behalf of simulated
. .
. .
users. InetLoad has a graphical user interface (GUI) that lets you set the number of users it is
. .
simulating and the number of worker threads that cycle through the script on behalf of each
. .
. .
simulated user. Experiment with setting the number of users simulated and the number of worker
. .
threads. This will help to avoid response time artifacts based on the client system's performance
. .
. .
. .
. .
In order to simulate large numbers of users, one or more copies of InetLoad must be run manually
. .
on multiple client systems. Each copy of InetLoad produces its own log file containing command
. .
. .
response time information. To determine the response times per command type from multiple
. .
InetLoad logs, it is necessary to use Islog, another Microsoft tool, that merges and produces
. .
. .
statistics from log files.
. .
. .
. .
. .
. .
. .
. .
. .
6
M
E A S U R E M E N T S