Check out the new USENIX Web site. next up previous
Next: Second experiment Up: Experimental results Previous: Measurements

First experiment

 

  figure127


Figure 1: Ideal file arrival times vs. estimated file arrival times for drawtest (experiment 1).

 

  figure135


Figure 2: Ideal file arrival times vs. estimated file arrival times for tictactoe (experiment 1).

 

  figure143


Figure 3: Ideal file arrival times vs. estimated file arrival times for Argo/UML (experiment 1).

 

  figure151


Figure 4: Ideal file arrival times vs. estimated file arrival times for Java2D (experiment 1).

 

  figure159


Figure 5: Ideal file arrival times vs. estimated file arrival times for HiNote (experiment 1).

 

  figure167


Figure 6: Download sizes for zlib bundles and Pack bundles vs. cumulative zip and Pack (experiment 1).

In the first experiment, bundlings of the rt.jar subset were generated using a collection of 17 input profiles, from the following applets and applications:

We evaluated the bundlings produced from the input profiles on five applications: the Drawtest applet, the TicTacToe applet, Argo/UML, the Java2D demos, and HiNote. For all five applications, the profile used in the experiment was a member of the input profiles used to generate the bundlings.

Figures 1-5 show the expected file arrival times vs. ideal file arrival times for the applications. In general, the bundlings were competitive with the ideal arrival times, although the `strict' bundling 1.0-200-5 fared noticably worse than the `loose' bundlings. This shows that even when network latency is relatively low (4 ms) it is still an important factor.

The vertical `cliffs' in the plots correspond to files being sent earlier than when they are needed; essentially, they delay the file whose request is currently outstanding. The diagonal sections (where the slope diverges from the ideal) corresponds to request latency. The horizontal `plateaus' in the plots correspond to files already received being used by the client. These satisfy their requests instantaneously, since the data is already available.

Figure 6 shows the download sizes for the applications and bundlings in experiment 1, for bundles in both zlib and Pack formats. All of the zlib bundlings were competitive with cumulative zip, although they were much larger than an equivalent Pack archive would have been. The `loose' Pack bundlings (0.8-1000-200 and 0.8-1000-500) were able to get quite close to the compression ratio of a single Pack archive, while the `strict' Pack bundling (1.0-200-5) had somewhat lower compression than a single Pack archive. All of the Pack bundlings were significantly better than cumulative zip.


next up previous
Next: Second experiment Up: Experimental results Previous: Measurements

David Hovemeyer
Tue Feb 27 18:43:09 EST 2001