Adaptive algorithm selection method (AASM) for dynamic software tuning

This paper presents a new approach to dynamic software tuning called the adaptive algorithm selection method (AASM). The AASM is built into the calling sequence of a library. When the library is called, the AASM is activated. The AASM selects and executes the optimum algorithm from registered algorithms in a library, based on data and machine type. As a result, the software is automatically tuned and the execution time is shortened. The relation between the data and the best algorithm for a given machine is learned by a neural network from the results of performance tests of the registered algorithms. We experimented on a multi-strings search problem with the AASM on the following machines: the CRAY X-MP/216, FACOM M 1800/30, and SUN Sparc Station 2. From these experiments we demonstrated that the AASM is able to minimize the execution time.<
>