Preliminary Results

 We have used the Bayanihan framework to write a simple application that factors a Java long integer N by dividing the search space $\{1,...,\sqrt{N}\}$ into fixed-sized work packets which are executed by the work engines. This problem, inspired by [41], although somewhat unrealistic, was chosen because its simplicity and predictability make it easy to measure performance and analyze results.

Figure2 shows some results for N=1234567890123456789, with 112 work packets of size 100000000 each. We used five identical 200MHz dual-Pentium machines (one server, and four clients) connected to a 10Base-T Ethernet hub, and running Windows NT 4.0. The server application was run using Sun's JDK 1.1.4, and the worker applets were run using Netscape 4.02. The speeds in the figure represent the average rate (i.e., numbers checked per millisecond) at which work packets were processed, over the course of searching the whole target space. The number of workers represents the number of worker applets run on the client machines. Each client machine was used to run up to two worker applets. The ideal speed is computed as the pure computation speed (measured at the client side) multiplied by the number of workers. With eight workers, we get the figures in Table1.

  

Figure 2: Timing measurements

  

measurement num/ms

pure computation (1 worker) 1828

ave. speed (8 workers) 13571

speedup 7.42

efficiency 92.8%

To evaluate the performance of Java, relative to C, we wrote a simple C program, compiled it with djgpp [42] (the gcc implementation for MS-DOS). Comparing its performance with the one-worker pure computation performance, we get the results shown in Table2. These results show that at least for simple tight-loop computations such as our factoring algorithm, the performance of the Java just-in-time-compiler in Netscape is comparable to native C code.

  

version num/ms

C 1879

Java 1828

While these results are admittedly preliminary and based on a toy problem, we believe that they give us hope that Java-based implementations of volunteer computing will be feasible and practical.