Throughout this chapter we have been using an interpretation penalty of 10 for our continuous compiler. While this appears to be a reasonable value to use in light of recent work [19], a quick look at the performance of the continuous compiler with other interpretation penalties would be worthwhile.
Table 6.24: Performance with Various Interpretation
Penalties
Table 6.24 shows the performance of the continuous compiler for ghostview, pico, and render with interpretation penalties of 3, 10, and 100. The longest-overall compilation strategy and the replace-preemptive model were used.
As expected, the continuous compiler performs better with a smaller interpretation penalty and does worse with a larger penalty. The surprise is that the variation in performance is so small. Even when the interpretation penalty is increased by a factor of 10, from 10 times to 100 times, the slowdown of render is only about 3%. With ghostview the slowdown jumps to about 44%, but that is still remarkable considering that it corresponds to an increase of 900% for the interpretation penalty.