Storage /

How we did it

By Edwin Mier and Kenneth Percy
Network World, 08/21/00

The storage-area network test bed used the same traffic sources - seven Windows NT servers - the same Fibre Channel adapters (QLogic model QLA2200F/33) in the servers, and the same target disk subsystems for testing the SAN switches. That way, the only variable was the performance difference of the SAN switches under test.

SANitycheck
Building an enterprise SAN
Interactive scorecard and NetResults: SAN Switches
Archive of Network World reviews
Subscribe to the Product Review e-mail newsletter

Except for Gadzoox, the switch vendors each submitted four switches each, which we interconnected in a diamond topology. Gadzoox only provided one switch. We interconnected the switches to test the ability of the SAN switch fabric to failover and reroute traffic if a switch or an inter-switch link (ISL) fails. We also tested each vendor's switch in a single-switch environment, with servers and disk storage connected via the same switch. Gadzoox was not yet shipping support for multiswitch topologies at the time we tested, so testing with that vendor involved just a single-switch topology. While Gadzoox was reportedly beta-testing a new Fabric Switch Module, it was not available for our testing.

To generate traffic - in our case disk I/Os - we used one to seven servers running NT 4.0 with Service Pack 6a. These were identical Pentium III, 500-MHz machines with 128M bytes of RAM. The SAN server interfaces, or Host Bus Adapters (HBA), were adapters, each providing a Fibre Channel, short-wavelength fiber-optic connector, and using the same drivers. The switch vendors were advised in advance of the testing, and accepted our decision to use the QLogic HBAs for the tests.

We installed a freeware software test tool from Intel, called IOMeter, on the servers. IOMeter is a workload generator and measurement tool that can execute disk I/O read and write loads, and meticulously report the performance results. This software tool lets one server station act as a master, controlling the set-up and execution of the tests with the other servers, then collecting and consolidating the performance data. We used IOMeter Version 1999.10.20.

For our I/O disk subsystem targets, we used two of Eurologic Systems' Eurologic XL-400 JBOD systems, each with seven 18G-byte Seagate Cheetah 18LP disk drives and a native Fibre Channel interface. The two JBODs were cascaded, yielding a total of 14 possible disk I/O targets.

To confirm the throughput performance and measure latency, we employed the Gigabit Traffic Analyzer (with a 256M-byte trace buffer) from Finisar.

For throughput, we tested one server initiator to four target disks - first through one switch, then through a multiswitch fabric. We then tested seven server initiators to 14 target disks again through one switch, and then through the multiswitch fabric. In each of the tests, we used the IOMeter to generate 10M-byte reads, then writes, then a combination of 50% reads and 50% writes. At least three iterations of each test were performed, and for each test we recorded total I/Os per second (how many times the 10M-byte file could be written or read per second), total megabytes per second of throughput, and the average I/O response times (elapsed time to complete a read or write, in milliseconds).

To measure latency, the Finisar Gigabit Traffic Analyzer time-stamped the first 10 SCSI commands issued by the server into the multiswitch SAN fabric, and then compared those with the time-stamp of the commands exiting the switch fabric. The average delta of the two times produced the latency measurement.

To measure failover time, we used IOMeter to generate 2K-byte random reads, from one initiator (server) to four disk targets, in a continuous stream. We then ascertained the active ISL and disconnected it. Then we launched random 10M-byte disk reads from seven initiators to 14 disk targets, and ascertained and disconnected the active link. For each test, the Finisar Gigabit Traffic Analyzer recorded the elapsed time from when traffic ceased to when it resumed.

We also ran some comparative tests of SAN performance involving NT-server back-up operations over a SAN, to a target tape system, rather than to disk drives.

Back to the main review