I’m delighted to release the first full and non-beta version of Stibium, my free benchmarking utility for storage performance testing, most particularly with SSDs. This now includes improved testing, greatly improved statistical analysis, an extensive Help reference, and an improved interface.
The Help reference amounts to some 21 pages, with 5,000 words including excerpts of relevant source code and details of the statistical methods used. This is only the first version of this reference, which will expand further in the next version. Even if you’re familiar with Stibium, please take a little time to browse this. As usual, I provide a separate copy of the PDF document, as well as the one embedded in the app.
This version can now perform multiple write tests of fixed size files. Let’s say you’re particularly interested in measuring the performance of an SSD when writing and reading files of 3 GB. You can now generate multiple files of just that size, and measure that write and read performance.
Whenever you open Stibium, it displays the last read and write performance results you obtained in testing. This makes it easier to perform tests after a restart, although if you want to save the full report, you’ll need to export it as a text file, or copy and paste it, before quitting the app.
I’ve laid out its controls afresh and relabelled them to make their purpose clearer. This should make it easier to use for novice and expert alike.
In the analyses performed automatically, Stibium still calculates and displays median values for smaller groups of measurements, but where numbers allow it now gives 20% trimmed means instead. As an estimate of a measurement, these are more robust and better estimators. They’re produced by removing the largest and smallest observations, then calculating the average of the remaining central values.
In addition to performing linear regression, whenever it can, Stibium now works out an overall transfer rate – with a confidence interval – using the Theil-Sen method. In this, it calculates the gradient of every possible line fitted through each pair of points in the dataset, then works out the median of all those gradients, which it converts into a transfer rate. This is more resistant to the effects of outliers than least-squares linear regression, although it’s quite computationally intensive.
Finally, the new Help reference provides a step-by-step guide to performing the ‘Gold Standard’ test, and details the main types of tests available.
I look forward to your comments and suggestions. I have major commitments through the end of the month, but early in February intend adding its graphical displays of results.