PC under Unix as a Web Server

PC under Unix as a Web Server

Load and traffic simulation of a PC under Unix to be used as a Web server.
Profile of the new Web server.

Configuration used for testing

Web Server: PC Pentium I, 100MHz, 64MB RAM, EDI hard disk, 10 Mbit Ethernet network connection, Solaris 2.5 system, NSCA/Purl Web server
Web Client: Sun SPARC 20, 150 MHz
Network: Both server and client are connected to the same sub-segment on an Ethernet Switch
Simulation program: Developped with Perl
Simulates 10 simultaneous clients for the first 10 minutes, then 15 clients for 10 minutes, etc. until it reaches 100 clients.
One client will request for 2 to 20 files (typically one HTML file and the associated pictures) the pause for 5 to 120 seconds (time needed to read the page before the user jumps to another page).
The URL are obtained from the list of URL that were acceessed for the server in real life during the most 2 busy days.

Simulation 1

The first three graphs represent the usage of the Web Server during the simulation.

Fig. 1: CPU usage and average

On the figure 1, it can be noticed that the CPU average load stop growing after 50 or 55 clients, this is due to an overload on the machine used to simultae the clients. The client machine could not simulate more than 50 clients realistically.

For 50 simultatenous simulated clients, the CPU load is around 60%, the machine can handle it pretty well.

Fig. 2: Disk accesses and average

Fig. 3: Memory usage and average

Figure 2 and 3 show the disk access and page swap respectively. Both are far bellow the maximum that the machine can handle (disk high load is 300 accesses per second).

The number of page fault is constant when the number of clients increases, so the memory size is not a slowing factor.

From the client point of view we get the following statistics.

Fig. 4: Number of second to transfer one file

A small increase can be noted around 50, 55 clients, this once again is due to the lack of speed of the machine used to simulate the client. The average transfer time, at 2 or 3 seconds per file is satisfactory.

The figure 5 represents the transfer rate in bytes per second.

Fig. 5: Number of second to transfer one file

Simulation 2

In order to check the influence of the internal disk cache of the server, the simulation was run from the list of files on the server instead of the most 2 busy days in the server logs.

Fig. 6: Disk accesses and average

Disk usage is comparable to the one in simulation 1 and far below the limit of the machine.

Conclusion

The simulation shows that a Pentium I, 100MHz machine, with 64 MB RAM can handle the load of up to 50 simultaneous Web clients. It has provision of disk speed, CPU ressource and the memory is not over loaded.

To make a comparison, one user loading 2 to 20 files every 5 seconds to 2 minutes would load approximately 10 files per minute, that is 1,440 files per day. Fifty simultaneous users would load 72,000 files per day. On the busiest day of the server www.sat.ait.ac.th, only 1,057 files were requested.

The simulation was only considering plain HTML files, in order to handle compplex CGI, more and more SSI extension, link to databases, etc. some more resources will be needed.

I recommend a Pentium II, 266MHz or above, with 128 MB RAM.

As for the operating system I suggest Free BSD, version 2.7 that prooves to be stable and does not over load the system as much as Solaris.

The hard disk can be a Fast Wide SCSI 3 disk of 4.3 GB, this is enough to accomodate the operating system and several hierarchy of web sites. Typically, CSIM Web site is 160 MB and SAT site is 140 MB.

As a Web server, I prefer Apache, with extension for Secure Socket Layer (SSL) for secure transaction and suExec wrapper for security of the CGI scripts. This server will be able to accommodate the web sites for CSIM , SAT, SAT Electronic Journal through multi hosting facility, and we will be able to offer web hosting for other school or program of AIT as well as for some external institutions (AI3, TH-NIC, etc.) provinding a full range of services on a single management basis.