Performance - Write
If you think that having a gigabit Ethernet interface on a consumer-priced NAS guarantees high performance, the TeraStation has a few surprises in store for you. As I do with all NAS testing, I used IOzone to check out the TeraStation's file system performance (the full testing setup and methodology are described on this page).
I tested the TeraStation in each of its four possible modes: Standard (250GB×4); Spanning (1TB×1); Mirroring (RAID 1) (250GB×2); and Parity (RAID 5) (750GB×1). For each mode, I ran the test with the Terastation and computer running iozone connected via a 10/100 switch, then moved them to a NETGEAR GS108 10/100/1000Mbps switch for gigabit-speed testing.
The iozone computer connected via an Intel PRO/1000 MT Desktop gigabit Ethernet adapter, which was set to auto-negotiate its connection speed. Note that since the GS108 switch doesn't support Jumbo Frames, I couldn't test its effect on the TeraStation's performance.
• Firmware version tested was 1.34
• Keep in mind that the maximum raw data rate for 100Mbps Ethernet is 12500 kBytes/sec and 125000 kBytes/sec for gigabit
Figure 7 shows a 3D surface plot of the TeraStation write performance using file sizes from 64 kBytes to 128MBytes and record sizes from 4 kBytes to 16MBytes. Note that the plot is oriented so that larger file sizes are closer to the front.
Figure 13: TeraStation Write performance - 100Mbps Span mode (default)
(click image to enlarge)
This plot is typical of those that iozone produces, with the peak speeds that are well above Ethernet speeds produced due to both Windows OS and NAS caching and buffering. But when I compared it to an iozone run performed under the same test conditions on a 160GB LinkStation, I was surprised by the result shown in Figure 14.
Yes, those are the same vertical scales and, yes I re-ran the tests to make sure they were repeatable. The LinkStation uses a 200MHz PowerPC processor (vs. the TeraStation's 266MHz PowerPC), and has 64MB of RAM (vs. 128MB for the TeraStation). But the two products use different hard drives and IDE controllers, so the results aren't unexpected - at least that's what a Buffalo applications engineer felt when I shared these results pre-review.