Courage in profiles

Abstract

This paper challenges the myth that Lisp programs are slow, and C programs are fast, by comparing two 2,000 line programs. The two programs are popular machine learning programs, Fahlman's Cascade Correlation program, CASCOR.1, and John Koza's Genetic Programming program, referred to here as GP. CASCOR1 is floating point intensive, while GP is floating point and structure manipulation intensive. These programs not only provide medium sized interlanguage benchmarks, but also provide an opportunity to study performance tuning. One man-day of tuning was done on each of the Lisp versions of the programs, using commonly available tools such as compiler advice and profilers. Much less than 10% of the code was changed. The performance of CASCOP-I (already highly optimized) was improved by almost 40%, comparable to or slightly faster than the C version. The main difference between Lisp and C is the quality of compiling a dot product loop. GP (not optimized, but well written pedagogical software) was improved by a factor of 30, and is over twice as fast as the C version which was advertised as "highly optimized". Lisp's advantages are due, at least in part, to automatic garbage collection and dynamic typing. In real programs, performance has more to do with choice of algorithm, and quality of compiler than on choice of language.