Issue TOC Print full article Print this page Download as PDF E-Mail this URL Discuss this article Flame the author     In this article Introduction Dell Computer Corp. Hewlett-Packard Co. RLX Technologies How We Tested Savantis Introduces Its DAN Online Only: Sweet Box, Big Price

We performed several system tests to analyze the servers' disk I/O and Web performance. Intel's Iometer 2003.02.15 (sourceforge.net/projects/iometer) provided a performance validation of the hard disk subsystems by generating I/O workload and performing a variety of simulated tasks. Spirent Communications WebAvalanche tests provided raw Web performance numbers under an untuned copy of Microsoft IIS.

Iometer Tests

We tested with three basic Iometer tests: database, maximum throughput and maximum I/O rate. The database test used a 2-KB transfer request size, with a mix of 67 percent reads and 33 percent writes. This proportion represents a typical database workload.

On the maximum throughput test, we used a 64-Kbps transfer request, and on the maximum I/O rate test, a 512-byte transfer request. On both tests we set the read/write distribution to 100 percent read and the random/sequential distribution to 100 percent sequential.

Intel Iometer Max I/O Test click to enlarge

We also customized a test, which we performed three ways: with a 512-byte transfer request size and distributions set at 100 percent read and 100 percent sequential; with a 2-KB transfer request with a 67 percent read, 33 percent write distribution and 100 percent random; and with a 64-KB request set back to 100 percent read and 100 percent sequential.

Spirent WebAvalanche

We used Spirent's WebAvalanche 4.0 to measure transactions per second under Microsoft IIS, untuned (see results in chart, left). We set up this test with a 10-second ramp-up time, 100 users per second. File sizes were set at 7 KB to 10 KB. If the server blade had a 10/100/1000 network connection, a single adapter was used. With multiple 10/100 network connections, adapters were teamed under a load-balanced configuration to provide increased performance. Thus we were almost always processor-limited in our tests, rather than network I/O-limited.