The Tak Benchmark Applet

Deep recursion can test the speed with which a language can make method calls. This is important because modern applications have a tendency to spend much of their time calling various API functions. PCWeek invented a benchmark called Tak which performs 63,609 recursive method calls per pass. The algorithm is simple: If y is greater than or equal to x, Tak(x, y, z) is z. This is the nonrecursive stopping condition. Otherwise, if y is less than x, Tak(x, y, z) is Tak(Tak(x-1, y, z), Tak(y-1, z, x), Tak(z-1, x, y)). The Tak benchmark calculates Tak(18, 12, 6) between 100 and 10000 times and reports the number of passes per second. For more information about the Tak benchmark see Peter Coffee's article, "Tak test stands the test of time" on p. 91 of the 9-30-1996 PCWeek. (The article may be on PCWeek's web site somewhere, but regrettably that site, while it looks pretty, is lacking some basic navigation aids. I was unable to locate the article, either directly or through their search engine. If anyone finds the URL let me know, and I'll report it in the next issue of this newsletter.)

The applet below is my variation of this benchmark. You can select the integer or floating point versions of the Tak method and specify the number of passes to be made through the benchmark. The java.util.Date class times that part of the test where the benchmarking is done.

If Java were turned on in your browser, you'd see the applet now.

My Powerbook 5300 achieved speeds between 3.5 and 5 passes per second on this test. Sun's Mac VM was about 10% faster on this test than Natural Intelligence's. The heavily loaded Sparcstation at sunsite.unc.edu (load average 4+) achieved a little more than 3 passes per second. Given the various external factors affecting machine performance, these are hardly scientific measurements. I'd be curious to hear what your results are.

The Applet Source
The Benchmark Source