8. Performance - Comparative Tests

To get a better feel for the non-buffered performance, it's more helpful to look at the tests for a 128 MByte file size. Figures 14 and 15 show write and read performance of the SC101 for 128 MByte tile sizes compared to an assortment of high-performance NAS products, plus the iozone machine hard drive. The products compared are the Buffalo LinkStation, Buffalo LinkStation Gigabit, Buffalo TeraStation and Infrant ReadyNAS 600. The TeraStation's results are with it configured as four separate drives (no RAID), and the ReadyNAS' data is from tests run with it configured for RAID 5 operation.

For write, the SC101's runs with the pack with no mirroring, but turns in the lowest performance with mirroring, with about a 50% throughput drop. Best write performance comes from the iozone machine's hard drive, followed by the Infrant ReadyNAS 600.

Figure 14: SC101 comparative write performance
(click image to enlarge)

The read plot is a bit odd, since if I had set the vertical scale to show the hard drive's results, you wouldn't have been able to see much for the other products. So the hard drive's results don't appear on the plot. The two lines that dive in from off-scale at the 256 kByte record size point represent the SC101's mirrored and non-mirrored runs. The non-mirrored SC101 results are still a bit better than the mirrored, and don't have the 50% performance gap that writing has.

Figure 15: SC101 comparative read performance
(click image to enlarge)

Since I've had readers tell me that it's difficult to sift through all this data, I've decided to add a file copy test to try to simplify things a bit. The reference setup is just two systems running WinXP SP2 - one an Athlon 64 3000+ w/ 512 MB, the other a 2.4 GHz Pentium 4 w/ 504 MB - connected via a 10/100 switch.

I then just drag-and-drop copied a folder with 1.70 GB of data (1,831,575,112 according to WinXP) in both directions to exercise the target drive for both write and read, starting my stopwatch when I dropped the file and stopping it when the file copy window went away. The test folder held an assortment of file types (including zip, exe, wmv, txt and iso) with doubled sizes (i.e. 1 MB, 2 MB, 4 MB, etc.) from about 1 MB to 700 MB.

I ran each test once for the "reference" networked Windows drive, and 20 GB SC101 non-mirrored and mirrored drives, with the results summarized in Table 1.

These results are pretty much in line with my iozone observations, except for slightly better write performance when the Athlon machine wrote to the SC101 vs. its own internal IDE drive.