Experiments and Results

We ran 7200 experiments to assess the performance of our implementations. 3600 of them ran with the starfield display and the preview bar enabled and the other 3600 ran with them disabled. The following values were used in our experiments:

Number of attributes

(a)
2, 4, 6, 8, or 10

Starfield size

(f)
, , or pixels

Point size

(d)
, , , or pixels

Range slider size

(p)
150, 200, or 250 pixels

Dataset size

(r)
10,000, 25,000, 50,000, 75,000, or 100,000 records

Jump size/range slider size

(j/p)
1/50, 1/25, 1/10, or 1/5

In the first 3600 experiments we measured the time to update the internal data structures (without any user interface updates). In general we observed that the ``pure'' querying time is no more than 20 milliseconds (average of 10 milliseconds). This was negligible, with respect to the starfield display times obtained, especially when the number of records got bigger (with a starfield display that does not use the incremental approach this time becomes even less significant with respect to the update times for the starfield display which suggests that the starfield update times must be optimized first for a faster DQI).

For the complexity analysis the remaining 3600 experiments were run, which we present in the following subsections (Figure 5).