How We Did It

How we tested the midrange Windows-powered NAS devices.

To generate disk I/Os, we used 10 HP ProLiant ML350 servers – each with a minimum composition of two Pentium III/866-MHz processors and 1G byte of memory – and installed IOMeter on each server. IOMeter is an open-source workload generator that can execute user-definable disk I/O read and write loads, and report the performance results. We used IOMeter Version 2001.07.19, and it is available at http://sourceforge.net.

Each server was configured to emulate two discrete NAS clients, for a total of 20. A share from the NAS was mapped on each server to serve as the storage target. To service Read requests, IOMeter created a file on the target NAS that was 1.3 times the available system memory used for I/O caching on each NAS device tested. This was done to prevent caching of the entire file, which would prohibit adequate stressing of the device’s storage subsystem. Also, all devices tested were populated with four drives, so none had the performance advantages that come with striping across more disks.

Two separate client emulations were created, one for file server clients and one for Web server clients. The file server client emulation consisted of 4K-byte request sizes, 80% read, 20% write and 100% random access to the test file. The Web server client consisted of 100% read and 100% random access to the test file of various request sizes. Three 10-minute iterations of each client were performed, and results were reported as the average I/Os/sec and M byte/sec of the three.

In separate tests, three standard, garden-variety denial-of-service attacks, not custom-made DoS attacks, were launched against the NAS devices while IOMeter testing was in progress, again in 10-minute iterations. Degradation was reported as the average percentage decrease of I/Os per second recorded from the results of testing with no DoS attacks in progress.