M. Poldner and H. Kuchen (Germany)
Algorithmic Skeletons, parallelism, divide and conquer
Algorithmic skeletons intend to simplify parallel program ming by providing recurring forms of program structure as predefined components. We present a fully distributed task parallel skeleton for a very general class of divide and conquer algorithms for MIMD machines with distributed memory. This approach is compared to a simple master worker design. Based on experimental results for differ ent example applications such as Mergesort, the Karatsuba multiplication algorithm and Strassen’s algorithm for ma trix multiplication, we show that the distributed workpool enables good runtimes and in particular scalability. More over, we discuss some implementation aspects for the dis tributed skeleton, such as the underlying data structures and load balancing strategy, in detail.
Important Links:
Go Back
