|
|
3.1.2 Network Characterization of the Apache/Linux System
James J. Filliben
Debra Tang The Apache workstation in combination with the Linux operating system is a popular industrial choice for the processing of network traffic. To assist in this characterization, Debra Tang of the Advanced Network Technologies Division of ITL has developed and implemented an on-line interactive toolkit (APALIN) for the interactive monitoring and analysis of processing times from such a system. This toolkit makes use of the non-obstructive NIST-developed multikron chip which measures and stores elapsed times for various internal function executions, and does so with a minimum of interference. The primary response variable for the system is total processing time in the handling of a net request. This total processing time may be partitioned into 4 major tasks: 1. accept_request; 2. read_request; 3. perform_request; and 4. disconnect_request. These 4 tasks have a variety of sub-tasks (example: parse_url, check_hostalias, log_transaction, unescape_url, translate_name, directory_walk, header_parse, invoke_handler, etc.)
Natural questions which arise in the characterization and analysis
of the Apache/Linux system are, for example,
1. Where does the system spend its time? A full experimental characterization of the Apache/Linux system would take approximately 15 experiments. An initial set of 4 experiments were in fact carried out: 1. assess system stability and reproducibility; 2. determine the effect of stress on system performance; 3. determine the effect of drop rate, band width and file size on system performance; and 4. determine the effect of drop rate, bandwidth, delay time, file size, and file type on system performance.
The above problem setup lends itself nicely to the application
of standard experiment design construction and analysis techniques.
A series of designed experiments were run. In particular, a 2**3 full factorial was
run for experiment 3 above, and a 2**(5-1) fractional factorial for
experiment 4 above.
From these designed experiments,
factor rankings, optimal settings, and a mathematical model were determined.
Figure 2: The interaction effects matrix has main effects on the diagonal and 2-term interactions off-diagonal. This matrix is for the analysis of a 2**(5-1) fractional factorial design. Interpretationally, it is seen that the dominant main effect is X4: file size, and the dominant 2-term interaction is X1*X3: drop rate x delay.
Date created: 7/20/2001 |