lients were divided into two groups, four systems conn
ected to each of the server's 100-Mbps backbones.
Three systems ran standard NT Server, while three ran different versions of Unix (Irix, Solaris, and OS/F). To level the playing field, we employed Netscape Enterprise Server 2.01 as the core Web software. FTP server software was the exception; the standard OS release was used. Netscape LiveWire Pro, which includes the Informix database system, sat on top of Enterprise Server. Netscape Mail Server handled all mail service.
Installation defaults were left largely unchanged. We made minor adjustments to tune the system for database or Web service activities as appropriate. We configured Informix shared memory to provide a reasonable balance of allocation of primary memory between Informix and other OS-related activities. Likewise, we modified processor and thread limits for each of the two test scenarios.
In the first scenario, we examined server performance under a
Web load
. Each of the eight clients exec
uted from one to eight concurrent tasks. Each task in turn requested a series of URLs (95 percent HTTP, 5 percent FTP) from the server. The requests were made into FTP and HTTP trees encompassing several thousand files that filled over 200 MB of disk space. The entire tree was cached in system memory -- a situation not uncommon in large server installations.
The Web test determined the server's ability to create and break network links, search cached file data, and transport information from memory to the network interfaces. We use standard "keep-alives" to simulate realistic browser interaction patterns. During all tests, a moderate, nonmetered mail processing load was placed on the server. This load represents common background noise and external server activities.
Our second scenario treated the servers as back-end
database
servers fielding complex queries into a midsize, highly cacheable database. A single query sorted, indexed, and summed a one-million-record database.
We configured all four processors in parallel, allowing Informix to dominate the system. We ran the test with one, two, four, eight, and 12 concurrently initiated queries. Ideally, execution times for one, two, and four queries should be identical. Any increase in time represents OS or system overhead.
illustration_link (14 Kbytes)
The time required to serve the URL requests increased linearly as more load was added to the system.
illustration_link (15 Kbytes)
Service is speedy when the systems are running up to four qeries. Beyond that, time increases proportionally.
Test Specs
