In case anyone's curious, the solution turned out to be making that script a bit more general so it can handle any number of time series on STDIN, and limiting the number of datapoints plotted to 900, giving 15 minutes of back history on the graph.

Once more than 900 data points have arrived, the file is rewritten every second to hold only the last 900 points.

That keeps the performance costs pretty reasonable; once all 15 minutes are displayed, the CPU cost on a dual-processor Linux box is about 1% CPU usage per series.
--
Mike

Edit: 2003-01-21: It turns out that setting

gnuplot.raise: False
[download]

in .Xresources makes gnuplot behave MUCH better; it stops popping up in front of other windows with each data point.